Because I live in Italy and earn in euros, the euro is the base currency of my portfolio, even though 94% of my investments are denominated in U.S. or Canadian dollars. I record each purchase and sale at its euro value on the transaction date, using the prevailing exchange rate, and then compare that historical euro cost to the current market value translated back into euro. As a result, my reported IRR captures both the performance of the underlying holdings and the effect of currency movements.

In 2025, the euro strengthened against both the U.S. dollar and the Canadian dollar, creating a meaningful headwind for my reported returns.

A stronger euro reduces the euro-denominated value of foreign holdings even when those holdings perform well in local currency. On an FX-neutral basis, excluding that currency effect, my 2025 MWRR was 18.8%, compared with 22.9% for the USD-denominated MSCI All Country World Index (ACWI)2, which I view as the appropriate benchmark for a globally invested portfolio. Put differently, the underlying portfolio performance was materially stronger than the headline euro-denominated figure suggests, even if it still lagged my primary benchmark.

In 2024, I moved away from funds and from positions that no longer fit my framework, restructuring the portfolio around individual businesses I understood better and was willing to own more decisively. In 2025, I reinforced that discipline. The portfolio became less a collection of interesting stocks and more a concentrated group of businesses I believe can compound intrinsic value at high rates over many years. Put differently, 2024 was about transformation; 2025 was about sedimentation.

Portfolio Turnover: Fewer Positions, Higher Conviction

Portfolio turnover in 2025 was meaningful, and it was driven by selectivity rather than activity for its own sake. I began the year with a portfolio spread across nine holdings (Amazon, Brookfield, North American Construction Group, Alphabet, MercadoLibre, Pinduoduo, Evolution AB, goeasy, and Visa).

By year-end, the portfolio was concentrated around eight higher-conviction businesses. Alphabet had grown from 10% to become my largest position at 23%, followed by Amazon at 20.2% (vs 22.1%), Brookfield at 18.8% (vs 17.8%), MercadoLibre at 12.9% (vs 9.9%), while new positions in Constellation Software and Uber had grown to 7.5% and 6% respectively.

That shift reflected a deliberate high-grading process. During the year, I exited Visa and North American Construction Group in the March-April selloff to fund higher-conviction businesses, sold Pinduoduo in November for portfolio-quality reasons, and began building positions in Uber and Constellation Software when dislocations gave me attractive entry points. Evolution AB remained in the portfolio at year-end, but at a 3% smaller weight than where it started, before I ultimately exited it in February 2026. goeasy also remained in the portfolio at year-end at a similar weight, before I exited it in March 2026. The result was a portfolio with fewer holdings, larger position sizes in my best ideas, and a clearer tilt toward the businesses I believe offer the best combination of quality, reinvestment runway, and underwriting clarity.

Portfolio Holdings’ Returns

Before turning to the qualitative progress of each business, it is worth looking briefly at how my holdings performed in 2025 on a position-by-position basis. In this section, I show the IRR of each position from December 31, 2024 to December 31, 2025, excluding the effect of foreign exchange, to isolate the underlying investment performance of each holding. Alongside each return, I also show its approximate contribution or detraction to portfolio performance, based on its weight at the start of the year.

I am only showing positions that I held at both the start and the end of the year. That keeps the comparison cleaner and more comparable across holdings, since a full-year IRR is more meaningful when measured over the same time window. Positions that I initiated or exited during the year are better discussed in the sections on portfolio turnover and exited holdings, where the context around timing and capital reallocation matters more than a simple one-year return figure.

Returns show what the stock did over the year, while contribution and detraction show how much that move mattered at the portfolio level. The rest of the article is my attempt to answer the next question: whether the underlying business strengthened, whether intrinsic value likely increased, and whether my conviction became stronger or weaker as a result.

• Alphabet Inc. (): 86.8% (+8.68 pp)

• Amazon.com Inc. (): 9.38% (+2.06 pp)

• Brookfield Corp. (): 25.8% (+4.64 pp)

• MercadoLibre Inc. (): 13.1% (+1.31 pp)

• Evolution AB (EVO.ST): -22.1% (-1.99 pp)

• goeasy Ltd. (GSY.TO): -17.5% (-0.88 pp)

  Level-Headed Investing is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Alphabet (Google): AI Did Not Dilute The Moat; It Widened It

Alphabet Inc. () had one of the most important years in the portfolio because 2025 moved the debate from theory to evidence. I used the March-April 2025 market selloff, when broad U.S. tariffs sparked a market scare and pushed the stock to a more attractive valuation, to add to my Alphabet position at more attractive prices.

The company delivered strong business performance while showing, more convincingly than ever, that its long-standing advantages in data, distribution, infrastructure, and research can be translated into real AI product and financial momentum. Full-year revenue rose 15.1% to $402.8 billion, operating cash flow increased 31.5% to a record $164.7 billion. Within that, Google Services revenue grew 12.4% to $342.7 billion and Google Cloud revenue rose 35.8% to $58.7 billion, with Cloud profitability continuing to scale meaningfully from 14.1% in 2024 to 23.7% in 2025.

That progress also reinforced the core thesis I laid out in my December 2024 article on Google’s AI leadership: the market was underestimating how well Alphabet’s existing strengths in data, distribution, infrastructure, and research would translate into AI leadership at scale. For years, the market framed AI mainly as a threat to Google Search. In 2025, that narrative started to break down. AI Overviews scaled to more than two billion users across over 200 countries, becoming the most widely used AI product in the world by a wide margin, while helping drive additional query growth, especially among younger users. That matters because it suggests AI is not weakening Google’s core franchise, but reinforcing it. Search remained resilient, YouTube continued to benefit from both advertising and subscriptions momentum, and Cloud became a far more important profit engine than it was when I initiated the position in October 2022.

What stood out most was the combination of research leadership and improving product velocity. DeepMind, which Google acquired more than a decade ago, remains one of the company’s most important strategic assets, and the release of Gemini 3 further strengthened that position. More importantly, Google is getting faster at turning frontier research into live products across Search, YouTube, Cloud, and its broader app ecosystem. In YouTube, multimodal AI is already improving recommendations and creator tools. Across the company, Alphabet increasingly looks less like a collection of products and more like a full-stack AI platform operating across models, consumer surfaces, infrastructure, and monetization.

Google Cloud exited the year at a revenue run-rate above $70 billion, grew 48% in Q4, and reached 23.7% operating margins for the full year, a dramatic improvement from around break-even profitability in mid-2023. I continue to think Google’s full-stack approach is becoming a real differentiator here: proprietary TPUs, leading models, strong infrastructure, and improving scale economics are reinforcing one another. Cloud is no longer just promising optionality; it is now a major engine of value creation in its own right.

Alphabet delivered one of its strongest operating years in recent memory, and my thesis strengthened in 2025: AI did not erode Google’s moat; it widened it.

Amazon: A Full-Stack Platform Built for the Next Cycle

Amazon.com Inc. () was one of the clearest examples in my portfolio of intrinsic value increasing in 2025. I took advantage of the March-April 2025 selloff to add more to my position. At 17.5x operating cash flow (OCF), near the low end of its historical range, the stock looked mispriced for a business that owns two category-defining franchises: AWS, the $140 billion run-rate leading cloud hyperscaler, and Amazon’s unmatched e-commerce platform, supported by a logistics network fulfilling well over $700 billion in GMV annually. My view was that the market was underestimating both the resilience of Amazon’s model and the length of its growth runway.

The numbers were strong, but the bigger story was the shape of the progress. Full-year revenue reached $716.9 billion (+12% YoY), operating cash flow rose to $139.5 billion (+20% YoY), and AWS revenue grew 20% to $128.7 billion. More importantly, AWS growth accelerated from 17% in Q1 to 24% in Q4 despite ongoing capacity constraints, showing that demand remains ahead of supply. Amazon is not spending heavily to create demand; it is spending because demand is already there. I continue to believe AI, especially inference workloads, can materially extend AWS’s runway, and that today’s infrastructure build-out should generate attractive returns over time.

Just as important, Amazon looked increasingly like a full-stack infrastructure and computing company with commerce attached. The company kept investing across cloud, chips, AI tooling, logistics, consumer interfaces, and connectivity. Alexa+, Amazon Leo (“Project Kuiper”), and AWS’s expanding Trainium and Bedrock ecosystem all reinforced that direction. While reported results included some noise from charges and impairments, the broader picture was clear: Amazon was laying more rails for the next decade.

I also continue to think the market undervalues the retail business. It is not just a low-margin merchandise engine, but a scale-driven platform benefiting from logistics density, automation, faster delivery, and a fast-growing, high-margin advertising business. I believe retail margins can move meaningfully higher over time, supported by efficiency gains already visible in the steady decline in per-unit shipping costs.

Put simply, my thesis on Amazon broadened in 2025. This is no longer just a story of strong retail plus AWS, but of an integrated industrial and computing platform with multiple reinforcing engines of value creation. Even after the stock’s appreciation, I still view Amazon as attractively valued relative to the quality of the business and its medium-term earnings power.

Share

Brookfield Corporation: More Levers, More Earnings Power, Same Discipline

Brookfield Corp. () was another holding where my conviction strengthened in 2025, even if the progress was less flashy than in some of my technology-linked positions. I initiated my position in October 2023 and also took advantage of the March-April 2025 market sell-off to add more shares at attractive prices. That fit well with my broader approach during the year: using dislocations to add to high-quality businesses when the gap between price and intrinsic value widens. Brookfield has long stood out to me as an asset-rich, high-quality company with an exceptional capital allocation record, strong insider alignment, and ownership of assets that sit close to the backbone of the global economy. 2025 made the case clearer rather than more complicated.

In 2025, Brookfield reported total distributable earnings (Brookfield’s own free cash flow metric) of $6 billion for 2025, versus $6.3 billion in 2024, while DE before realizations, the cleaner measure of recurring earning power, rose 10.6% to a record $5.4 billion from $4.9 billion the year before. The company made broad-based progress across all three major engines of the business. Asset Management continued to scale impressively, with fee-bearing capital rising 12% to $603 billion, fee-related earnings up 22% to $3 billion, and fundraising reaching a standout $112 billion, supported by flagship strategies and growing traction in retail and wealth channels. Wealth Solutions also advanced meaningfully, with DE reaching $1.7 billion as annuity sales lifted insurance assets to $143 billion, underwriting profitability improved sharply in P&C, and Brookfield kept redeploying insurance float into higher-yielding Brookfield-managed strategies. Meanwhile, the Operating Businesses remained a solid cash generator, producing $1.6 billion of DE, supported by renewable power, infrastructure, and private equity, while real estate fundamentals improved through stronger occupancy and leasing activity.

What Investor Day reinforced for me is that Brookfield is a compounding platform with several underappreciated growth engines. One important engine is Brookfield Wealth Solutions, where the investment-led insurance model continues to scale and contribute a growing share of DE. More broadly, Brookfield’s current earnings still understate the earnings power embedded in the platform. As funds are monetized and capital is returned to investors, Brookfield should begin recognizing a much larger stream of realized carried interest on its own balance sheet, which is one of the clearest reasons management expects DE growth to accelerate from here. Management is targeting high-teens annual growth in operating DE over the medium term, and when combined with the expected step-up in realized carry, total DE growth could approach 25% annually over the next five years. If Brookfield comes anywhere close to that, the current valuation still looks very attractive despite the stock trading near highs.

Not every business line was equally strong in 2025. Real estate remained softer in places, and some of the headline earnings still masked the underlying progress. But that is also part of why I like Brookfield: it does not rely on one segment, one macro outcome, or one source of value creation. It has multiple levers - asset management, insurance, operating businesses, capital recycling, and balance-sheet opportunism, and they are overseen by managers who have earned a high degree of trust.

Put simply, my thesis on Brookfield became clearer in 2025. This is a business with several embedded growth engines, led by management that has repeatedly shown it can allocate capital well across cycles. Even after the stock’s strong move, I still believe Brookfield offers an attractive combination of quality, valuation support, and long-duration compounding potential.

MercadoLibre: Still Compounding, Still Investing, Still Playing Offense

MercadoLibre Inc. () was a clear example in my portfolio of a business still compounding at a remarkable rate while deliberately choosing to reinvest rather than maximize near-term margins. Full-year net revenues and financial income rose 39.1% to $28.9 billion, while operating income increased 21.7% to $3.2 billion despite heavy investment in shipping, fulfillment, cross-border commerce, and newer credit cohorts. The company did not slow down to protect optics. It leaned harder into the opportunity set.

That progress was visible across both of its core engines: Commerce revenue reached $16.3 billion (+34% YoY) and Fintech revenue rose to $12.6 billion (+46.2% YoY), underscoring how balanced and powerful the business has become. GMV grew 26.4% to $65 billion, TPM rose 41.3% to $277.8 billion, and acquiring TPV climbed 32.3% to $188.1 billion. By Q4, net revenues and financial income had accelerated to $8.8 billion, up 44.6% YoY, while management pointed to broad-based strength in Brazil and Mexico, rising buyer frequency, and continued market share gains across the platform. Just as importantly, this was not a one-off spike in growth. MercadoLibre has now delivered 28 consecutive quarters of 30%+ revenue growth, which speaks to how unusual the durability of the business really is.

What gives me the most confidence is that MercadoLibre’s moat is not built around one product, but around the interaction of several. It is the leading e-commerce, payments, credit, and increasingly advertising platform across Latin America, operating in markets that remain under-penetrated and still offer a long runway. Its commerce platform helps bring users and merchants into its financial ecosystem, while its fintech products expand access to payments, credit, and digital banking and, in turn, deepen engagement across commerce. That flywheel is difficult to replicate because competitors tend to be strong in only one layer. A pure e-commerce player does not have the same fintech engine. A pure fintech player does not have the same commerce platform, merchant relationships, logistics footprint, or customer purchase data. MercadoLibre has the full stack.

Scale is a major competitive advantage here. In e-commerce, the first player to achieve real scale in a market often gains durable advantages in logistics density, fulfillment efficiency, merchant breadth, customer trust, and data. Those advantages improve service, lower costs, and make it harder for newer entrants to compete without sacrificing economics. MercadoLibre has the first-mover advantage in Latin America. The stock sold off from its highs partly on renewed competition and macro concerns, but neither of those are new. MercadoLibre has been competing, gaining market share, and operating through volatile political and economic environments for over two decades. Those concerns look more like recurring background noise than a broken thesis.

The apparent margin pressure in 2025 was the result of conscious choices that should strengthen the moat over time: The company effectively traded 5-6 points of operating margin for scale, speed, and deeper fintech penetration. MercadoLibre is still finding plenty of high-return ways to deploy capital across commerce, logistics, payments, and lending. Just as importantly, it did so without compromising the balance sheet, generating roughly $1.48 billion in adjusted free cash flow during the year even while stepping up investment intensity.

2025 also brought an important governance milestone. MercadoLibre announced that founder and CEO Marcos Galperin would become Executive Chairman effective January 1, 2026, with Ariel Szarfsztejn stepping into the CEO role. Founder transitions can expose fragility in a business, but here the change looked orderly and confidence-inspiring. To me, that reinforced the idea that MercadoLibre has evolved beyond founder dependence into a deeper, more institutional compounding machine. The business still has a long runway, management is willing to invest aggressively where returns justify it, and the ecosystem moat across Commerce and Fintech keeps widening. Put simply, my thesis on MercadoLibre strengthened in 2025: this remains one of the highest-quality growth businesses in my portfolio, and one that still looks far from finished.

Share Level-Headed Investing

Constellation Software: A Correction Gave Me The Opening; The System Gave Me Conviction

Constellation Software Inc. (CSU.TO) had been on my watchlist for years, but its quality usually came with a premium valuation. I began building the position in October, during the deepest drawdown in the stock’s history, when AI-driven software panic and succession concerns finally gave me an entry price I found attractive. Part of the correction was understandable given the rich starting valuation (peak 35.8x P/FCF), but the deeper derating struck me as an opportunity rather than a warning.

My thesis was that CSU is best understood as a reinvestment system: resilient vertical-market software (VMS) businesses generating cash, that cash being redeployed into acquisitions, and a culture designed to adapt without breaking. I was not buying a generic software company, but a compounding machine whose edge lies in niche mission-critical products, sticky customer workflows, disciplined capital allocation, and an unusually hard-to-replicate acquisition network.

The full-year 2025 results reinforced that view. Revenue rose 15% to $11.6 billion, operating cash flow increased 24% to $2.7 billion, and free cash flow available to shareholders (FCFA2S) rose 14% to $1.7 billion. Net income was much noisier, falling 30% for the full year to $512 million, largely because of items such as the IRGA/TSS membership liability revaluation and the revaluation of the Asseco investment rather than any break in the underlying operating model. That is exactly why, with Constellation, I pay more attention to revenue, cash generation, acquisition cadence, and reinvestment discipline than to any single year’s reported earnings line.

Just as important, the strategic picture remained intact. Constellation kept deploying capital into acquisitions throughout 2025, and the broader logic of the model did not weaken. CSU operates in the part of software where trust, workflow depth, switching costs, and accountability matter far more than flashy interfaces or short-term AI headlines. AI may make code easier to generate, but it does not make operational consequences easier to bear. That is why I think the market has been too casual in lumping Constellation into a broad “software is dead” narrative. Even management’s own framing has been more measured than alarmist: AI may create some risks around barriers to entry over time, but it may also improve internal efficiency and help portfolio companies build better products faster. At this stage, I still view it as far more likely to be a manageable variable than a thesis-breaking disruption.

The other major development in 2025 was leadership transition. In September, founder Mark Leonard stepped down as President for health reasons and long-time executive and COO Mark Miller took over, a change that understandably unsettled the market. To me, that moment actually highlighted the strength of the model. Constellation was built to outlast any one person. Its real intellectual property is not a single product or centralized strategy deck, but a culture of incentive alignment, decentralized decision-making, acquisition discipline, and capital allocation rigor. The continued strong execution at Topicus, Lumine, and other parts of the broader ecosystem reinforces that point. If anything, fear around succession and AI may prove helpful at the margin by softening valuations across the small software ecosystem and improving future returns on capital deployment. Put simply, my thesis on Constellation strengthened in 2025: this remains one of the highest-quality reinvestment systems in public markets, and the correction gave me the chance to buy it at an attractive price.

Uber: The Platform Thesis Became Harder to Dismiss

I initiated a position in Uber Technologies Inc. () in May, taking advantage of a dislocation driven in part by misplaced market fears around autonomous vehicles (AVs). My view was that the market was anchoring to an older version of Uber: a ride-hailing company with fragile economics and uncertain long-term positioning. What I saw instead was a high-quality, capital-light platform with strong network effects, rising engagement, and durable earnings power. Over the course of 2025, that thesis became easier to defend. Uber is increasingly best understood as a “global royalty on movement”: a structurally advantaged platform that clips a small fee on a growing volume of rides, food delivery, groceries, retail, and local logistics without owning most of the underlying physical assets. Since I initiated the position, the stock has actually declined, reflecting the market’s continued skepticism around AVs.

The financial progress in 2025 reinforced my thesis. Trips rose 20.4% to 13.6 billion, gross bookings increased 18.9% to $193.5 billion, and revenue grew 18.3% to $52 billion. More importantly, operating leverage became much more visible: free cash flow climbed to $9.8 billion (+41.6% YoY). By year-end, Uber had 202 million monthly active platform consumers (+18.1% YoY) and was operating at a 15 billion annual trip run-rate. That combination of sustained volume growth, rising engagement, and sharply higher cash generation is exactly what a maturing platform should look like.

The part of the thesis that became more interesting during the year was autonomy. I continue to believe AV commercialization will be more gradual, more geographically constrained, and more partnership-driven than many market participants assume. Regulatory complexity, manufacturing bottlenecks, and the operational burden of deploying physical fleets all argue against a simple winner-take-all outcome. That is why I view Uber less as a business threatened by AVs and more as a beneficiary of them. In 2025, Uber launched Waymo rides exclusively on its app in Austin in March and Atlanta in June, and management said in Q1 that it was already operating at a 1.5 million annual run-rate of AV trips on its network. Uber also said its Austin Waymo vehicles were busier than more than 99% of human drivers in that market, and later noted high utilization in Atlanta as well. Those are important proof points because AV economics depend heavily on utilization, and Uber’s marketplace is built precisely to aggregate demand, reduce idle time, and improve asset productivity.

Just as important, Uber broadened its AV strategy well beyond Waymo. During 2025 it expanded or announced partnerships with a wide range of players, including Baidu, WeRide, Lucid and Nuro. That matters because I do not think the long-term value lies in guessing which AV stack wins everywhere. I think it lies in owning the demand layer, the routing layer, the network density, and the consumer relationship. As AV technology spreads, the value of the driver should diminish while the value of the marketplace should increase. Put simply, Uber looks less like a transportation operator and more like a “global royalty on movement”, with a long runway for both growth and margin expansion.

Subscribe now

Exited Positions: Selling to Strengthen The Portfolio

I exited three positions during the year: NOA, V and PDD.

I exited North American Construction Group () and Visa () during the March-April selloff to reallocate capital into higher-conviction holdings at attractive prices: Visa sale at a 30% gain was driven by opportunity cost and NOA sale at a 13% loss was driven by deteriorating fundamentals. When the market dislocation created the chance to add aggressively to businesses where my conviction was higher, I chose to simplify the portfolio and move capital accordingly.

The sale of Pinduoduo () in November at a 27% gain came from a different, though related, logic. I did not sell because I had turned negative on China. In fact, my broader view on China remained constructive. I sold because I wanted to further raise the quality bar of the portfolio and redeploy the proceeds into businesses I understood better and where my conviction was stronger. Taken together, these exits reflected the same principle that shaped the rest of 2025: a concentrated portfolio should keep evolving toward the businesses I believe are my best long-term capital allocation opportunities.

Exited Positions in Q1 2026

Since the beginning of 2026 I have exited two positions: EVO.ST and GSY.TO.

After Evolution AB (EVO.ST) reported Q4 2025, I exited it in mid-February 2026 at a 37% loss: the thesis that had been weakening for over a year finally fell below my bar for a concentrated holding. The issue was that Evolution’s path to reacceleration was becoming less visible and increasingly dependent on external normalization rather than controllable execution. Consecutive quarters of deteriorating fundamentals, ongoing operational issues, weaker channelization trends in Europe, unresolved cybercrime-related disruption in Asia, and a growing mix shift toward lower-leverage regions all made the investment case harder to underwrite with confidence. By the time I sold, it was clear to me that better alternatives existed elsewhere in the portfolio.

I also exited goeasy Ltd. (GSY.TO) in early March 2026 at a 67% loss after the company admitted that parts of its prior financial reporting could not be relied upon. In its March 10, 2026 operational update, goeasy disclosed a large incremental Q4 2025 charge off, withdrew its prior Q4 outlook and three-year forecast, said prior-period delinquency and staging disclosures would need to be revised, and acknowledged a historical reporting practice at the LendCare subsidiary under which certain customer payments were recorded as received even though they were heavily delinquent and, in some cases, were ultimately not collected. At that point, the investment thesis was broken. What had looked, for decades, like a subprime lender with a long track record of great operational performance became an uninvestable situation marked by damaged trust, unreliable historical numbers, covenant stress, and a far more uncertain path forward. I sold immediately because once confidence in management and in the integrity of the reported financials is gone, there is no place for the stock in a concentrated portfolio.

What 2025 Taught Me

2025 was the year my strategy became clearer. In 2024, I transformed the portfolio by moving away from funds and from positions that no longer fit my process. In 2025, I reinforced that shift by selling two weaker lower-conviction names, exiting another for portfolio-construction reasons, and concentrating further around higher-quality businesses. That is what high-grading the portfolio means in practice: raising the standard for what deserves a place and steadily shaping the portfolio around businesses whose intrinsic value I believe can compound through scale, reinvestment, and strong strategic positioning.

One thing 2025 reinforced for me is that my best decisions often come when the market is most uncomfortable. I learned that lesson years ago during the COVID-19 crash. Back then, I was still investing primarily through diversified funds, and when markets bottomed around March 23, 2020, I did not let the drawdown dictate my emotions. I doubled down. In 2025, during the March-April selloff, I applied that same mindset again. I did not want fear, headlines, or short-term price action to make decisions for me. I wanted volatility to be something I could use, not something I merely endured.

The year also reminded me that market narratives are often less nuanced than the businesses they try to explain. AI was framed as a threat to Google, hyperscaler CapEx was treated as a warning sign for Amazon, AVs were seen as a reason to doubt Uber, and software was dumped into one broad fear bucket that obscured what makes Constellation Software different. In each case, the more useful question was not what the market feared, but whether the underlying business was making real progress in the face of uncertainty.

2025 also reinforced that the strongest moats increasingly sit at the ecosystem level, not the product level. What draws me to businesses like Alphabet, Amazon, MercadoLibre, and Uber is not just that they offer strong products, but that they own broader systems of infrastructure, distribution, data, customer relationships, and adjacent profit pools that reinforce one another. That kind of positioning is harder to disrupt than the market often assumes, and it helps explain why these businesses continue to strengthen even when the narrative around them turns negative.

Another lesson is that portfolio construction is not just about finding attractive ideas. It is about ranking them honestly. Selling PDD while remaining constructive on China made that point very clear to me. A theme can remain intact while a specific stock no longer deserves scarce room in the portfolio. The same applies more broadly: every holding has to compete for capital, and some simply stop clearing the bar.

I was also reminded that capital allocation quality matters more than neat short-term optics. The businesses I grew more confident in during 2025 were not the ones maximizing near-term margins or smoothing reported earnings, but the ones willing to absorb accounting noise, heavier investment, or temporary margin pressure when those choices appeared economically rational. Over time, I care much more about whether management is creating value than whether the quarterly numbers look tidy.

And sometimes, a weakening thesis does not break all at once. It erodes. That was the lesson of Evolution AB. What began as a case for patience eventually became a case for redeployment, as the path to reacceleration became less visible and more dependent on external normalization rather than controllable execution. That is not a comfortable realization, but it is an important one.

If I had to reduce 2025 to one sentence, it would be this: I am becoming less interested in owning what merely looks attractive, and more focused on owning what I can underwrite with conviction through volatility. That shift began in 2024. In 2025, it became more deliberate, and more real.

Final Thoughts

2025 was not my strongest year in reported return terms. It was a meaningful year in a way that matters more to me: the portfolio became clearer, stronger, and more aligned with how I want to invest.

I added more aggressively when the market gave me better prices. I upgraded the portfolio into businesses where my conviction is higher. I sold positions that no longer cleared the bar and I was reminded, again, that volatility is only useful if it helps me own better businesses, not just feel worse while doing nothing.

That is the real progression from 2024 to 2025. Last year, I transformed the portfolio. This year, I clarified what it is supposed to be.

A concentrated portfolio is not a collection of interesting ideas. It is a ranking of conviction, renewed over and over again. In 2025, I raise that standard.

“We prioritize investments that are most synergistic with our ecosystem and either deepen existing moats or create new ones. All of our investments are required to demonstrate a clear path to positive direct contribution under conservative assumptions”

— MercadoLibre Q4 2025 Shareholder Letter

MercadoLibre, Inc. () posted another blockbuster quarter - Q4 2025 revenue up 44.6% to $8.8 billion and FY 2025 up 39%, but the stock fell 8% as margins compressed. Management explicitly chose to “buy” growth by reinvesting 5-6 points of operating margin into free shipping, logistics, credit cards, cross-border, and 1P, and the operating KPIs (frequency, logistics efficiency, Mercado Pago engagement, and disciplined credit metrics) suggest those investments are working. I bought the dip with MELI now 13% of my portfolio because I see temporary margin pressure as moat-widening, not a weakening business.

MercadoLibre: The Commerce + Fintech Flywheels

MercadoLibre is the leading e-commerce and fintech ecosystem in Latin America. Commerce compounds via selection → demand → seller supply → logistics density → faster delivery → higher conversion and repeat, with ads monetizing that engagement; Mercado Envios is “integral and crucial” to reducing friction and enabling fast delivery at competitive cost. Fintech (Mercado Pago) started at checkout and expanded into acquiring, accounts, savings/investments, and credit; the edge is data + distribution, using “repeated interactions with our ecosystem” to underwrite and offer credit contextually. Together, commerce drives fintech adoption, fintech boosts commerce frequency, and the data loop strengthens underwriting, reinforcing the moat.

The runway remains large; management notes “penetration is roughly half the level seen in the US, UK and China, and we see no structural reason why the region should not reach similar levels”.

Subscribe now

Q4 2025 and Year-End 2025: 5 Key Results That Matter

The Q4 2025 results had plenty of moving parts, but five data points captured my attention.

1. Elite Growth at Scale: Quarterly Acceleration on Top of a Strong Year

Q4 net revenues and financial income were $8.8 billion (+44.6% YoY), capping FY 2025 growth of 39% YoY with operating income growing 21.7% YoY, despite stepped-up investment intensity.

On the earnings call, management framed the quarter as “robust operating trends that reinforce the strength of the MercadoLibre ecosystem”, supported by “two primary growth drivers: the acceleration of our commerce business and the rapid adoption and structural expansion of our fintech services”.

The company has now grown the top line above 30% YoY for 28 consecutive quarters, underscoring the “effectiveness of the long-term investment and customer focus that we have within our ecosystem”.

2. Commerce: Frequency and Retention Are Improving, the Clearest Signal of Durability

The headline is easy to summarize: more buyers, buying more often, across more categories. In Q4, Gross Merchandise Volume (GMV)1 was $19.9 billion (+36.8% YoY), unique active buyers reached 83.2 million (+23.6% YoY). Items sold accelerated from 39% YoY growth in Q4 2024 to 43% YoY growth in Q4 2025.

Items sold per unique active buyer rose to 9 items (+15.4% YoY). This matters because frequency is the hidden engine of e-commerce compounding: higher frequency improves retention economics, creates more ad inventory, and increases logistics density (which then improves unit costs and delivery speed).

Management directly connects this behavioral change to their value proposition investments in Brazil. In the shareholder letter, they wrote: “Free shipping has been one of the most important pillars of our value proposition for a decade […] Extending free shipping in Brazil to items starting from R$19 reflects this conviction […] The impact on behavior is clear”. They added: “Since June, new buyers in Brazil have purchased more items across a larger number of categories with higher retention than cohorts that joined prior to the lower free shipping threshold”. In plain terms: MELI is buying habit formation today to harvest operating leverage tomorrow.

Advertising is the high-margin layer on top of this Commerce growth. Mercado Ads’ revenue grew 70% YoY in the quarter, and management noted penetration is still small versus its potential. As frequency rises, ads should contribute more to Commerce revenue and help offset shipping-related margin pressure over time.

3. Logistics Scaled and Unit Economics Improved

In Q4, same/next-day shipments reached 191 million (+29% YoY).

For the full year the “network absorbed a 41% YoY increase in volume during the year (nearly 500 million) additional shipments”. Management highlighted what I consider one of the most important proof points in the entire shareholder letter: “Unit shipping costs in local currency fell 11% YoY in Brazil in Q4’25”, while “maintaining the high service levels that have built trust over time”.

4. Fintech: Moving From Payments To a Full Financial Institution

Management is explicit about their ambition:

In Q4, fintech monthly active users (MAU) were 77.9 million (+27.3% YoY) - management highlighted that “MAU have been growing close to 30% for ten consecutive quarters, as our value proposition continues to improve”. Assets under management (AUM) reached $18.8 billion (+78% YoY, and a record +24.8% QoQ), and the credit portfolio ended at $12.5 billion (+90% YoY, 4x in three years).

MELI uses the marketplace checkout as a distribution engine for credit, and the real advantage is the underwriting loop behind it. Management says maintaining strong asset quality is a key enabler of growth, achieved by “combining technology with our marketplace’s high-quality audience of buyers”, where “data from repeated interactions with our ecosystem feeds into our underwriting models”. As a result, MELI has “millions of pre-approved users who are offered credit at the marketplace checkout (and elsewhere)” which was particularly effective in Q4 2025 as the company “issued almost 3 million credit cards by leveraging our marketplace’s peak season”. On the merchant side, the flywheel is similar: 33.8% of marketplace sellers are taking Mercado Crédito to fund inventory and growth, reinforcing selection, availability, and GMV back on the commerce platform.

5. Margins Compressed by Design, and Management Quantified the “Moat Spend”

The market’s issue with the quarter was straightforward: profitability didn’t keep up with revenue growth. Q4 operating income was $889 million (10.1% margin, -25.2% YoY) and net income was $559 million (6.4% margin, -39% YoY).

MELI quantified the operating margin compression and explained it as deliberate, early-stage investment. In the shareholder letter: “We estimate the combined impact of our strategic investments […] was equivalent to 5-6 percentage points of operating margin in Q4’25”. They list the major initiatives: “lower free shipping threshold in Brazil, CBT, 1P and the credit card”. Management made the philosophy unambiguous:

One line item worth calling out is “Bad Debt”, which is MercadoLibre’s provision for doubtful accounts (expected credit losses recognized through the P&L as the credit portfolio expands). Management addresses this directly in the shareholder letter:

In plain English, MELI grew credit aggressively, and that growth requires building loss reserves up front, so even if underlying credit performance is healthy, provisions can still weigh on operating margin in the near term.

Year-End “Platform Proof”: Two Engines, Broader Footprint, Strong Cash Generation Despite Reinvestment

Stepping back from the quarter, the full-year numbers validate that MELI is scaling both flywheels. FY 2025 net revenues and financial income were $28.9 billion (+39.1% YoY). Commerce generated $16.3 billion and fintech $12.6 billion. Operating expenses rose 39% YoY to $9.66 billion, driven mainly by sales and marketing (+46.9%) and provision for doubtful accounts (+66.4%), as MELI is leaning into growth levers without letting the core cost base (product and technology +17.3% and G&A +11.9%) balloon.

This cost structure also helps explain the margin narrative: gross margin (COGS) and operating margin (OpEx) pressure in 2025 was largely the deliberate cost of scaling long-term initiatives: expanded free shipping in Brazil, ramping 1P and fulfillment capacity, and early-stage investments in cross-border trade and credit. Those moves increased logistics subsidies and lowered shipping-related monetization (contributing to modest take-rate pressure), while 1P, CBT international fulfillment, and newer credit cohorts remain margin-dilutive during their build-out phase. Management framed the investment as a conscious trade-off to deepen engagement and widen the ecosystem moat as these initiatives scale toward maturity.

The ecosystem KPIs kept compounding: FY 2025 GMV was $65 billion (+26.4% YoY), Total Payment Volume (TPV) was $277.8 billion (+41.3% YoY), and Acquiring TPV was $188.1 billion (+32.3% YoY).

One of the cleanest ways to see MELI’s “platform proof” is to zoom out. Since 2012, Commerce and Fintech revenues have both compounded aggressively: Fintech even faster, while operating income has scaled alongside them. That long-term trajectory also highlights a structural shift in the mix: Fintech is no longer an add-on to Commerce, and it is approaching parity in revenue contribution, which meaningfully changes MELI’s earnings power profile. Commerce drives engagement and purchase frequency; fintech drives monetization depth and balance-sheet products.

Crucially, MELI did not fund this expansion by burning the balance sheet. Despite heavy reinvestment, the company generated approximately $1.48 billion in adjusted free cash flow2 in FY 2025. The ecosystem is internally generating capital while still scaling aggressively.

Equally important is the geographic breadth of that growth. Brazil remains the largest engine while Mexico and Argentina are no longer “supporting markets” - they are material contributors with accelerating adoption across commerce and fintech. The ecosystem is diversifying its revenue and profit pools across multiple countries, reducing reliance on any single macro backdrop while reinforcing network density in each local market.

Share Level-Headed Investing

Valuation: Is MercadoLibre Undervalued?

When a company is rapidly expanding a credit portfolio, conventional free cash flow gets distorted by balance-sheet swings, especially cash deployed into new loans. That’s why MELI reports adjusted free cash flow (AFCF) and why I start there. In FY 2025, operating cash flow (OCF) was $12.12 billion. After backing out movements tied to customer funds and regulatory restrictions ($4.99 billion), subtracting CapEx ($1.33 billion), and adjusting for the net cash deployed into the loan book ($6.66 billion) and the net proceeds from loans payable/other fintech liabilities ($2.35 billion), MELI reported AFCF of $1.48 billion.

To approximate normalized cash earnings power, I adjust AFCF for reinvestment into loan growth and estimated growth CapEx: $6.32 billion = $1.48 billion (AFCF) + $0.53 billion (40% of CapEx) + $6.66 billion (loan book growth) - $2.35 billion (net loan funding). This is not a forecast - if MELI slowed these investments, growth would likely slow too. It’s a useful way to frame underlying earning power during an investment-heavy phase.

On today’s ~$86.5 billion market cap, that implies ~13.7x price-to-normalized earnings power: modest for a business compounding revenue above 30% for 28 consecutive quarters while investing heavily into logistics density, fintech adoption, and credit penetration.

Final Thoughts

MELI exited 2025 with rare fundamentals: 28 straight quarters above 30% growth while investing aggressively in logistics, payments, and credit. The stock sold off on margin compression, but management was explicit that this was a deliberate choice tied to specific initiatives. I’m staying focused on the operating scorecard - frequency, logistics efficiency, fintech engagement, and credit quality - because that’s what will drive long-term earnings power.

By Q3 2025, my stance had shifted: those same issues were starting to look structural. Europe’s ring-fencing was becoming a lasting headwind, Asia’s counter-piracy measures were proving hard to stabilize without collateral damage to legitimate demand, and operating leverage had flipped from a tailwind to a constraint.

The Q4 2025 year-end report reinforced that framework. Europe is now constrained by declining channelization (players pushed to unlicensed operators Evolution won’t serve), Asia remains “slow and methodical” with no timeline, and incremental growth is shifting to regions with lower operating leverage. That combination reduced my confidence in a durable re-acceleration, so I exited after Q4.

What Q4 2025 Results Said

Evolution reported Q4 net revenues of €514.2 million (-3.7% YoY, +1.4% QoQ). Live Casino remains the engine: €438.6 million (85.3% of revenue) versus RNG at €75.7 million (14.7%). Live has effectively plateaued since early 2024 (€430-€460 million per quarter), while RNG has grown modestly (€70-€76m). The mix matters: without renewed Live growth, the consolidated top line struggles to compound.

Live revenue contracted 4.3% YoY in Q4.

From a cash perspective, operating cash flow (OCF) was €296.6 million in Q4 (-10% YoY), and management highlighted “Operating Cash Flow after Investments”1 of €258.2 million.

For 2025, total operating revenue was €2,119 million (-4.3% YoY). OCF was €1.25 billion (-3.5% YoY); less €134.8 million of CapEx implies €1.12 billion of free cash flow. The tension is clear: the company is still executing, but the revenue engine isn’t responding like it used to. As the year-end report puts it, “overall, we are proud but not happy with 2025”.

Subscribe now

Demand Is Maturing

From 2006 through 2023, Evolution benefited from two forces: strong execution and a steadily expanding global audience for online casino. The last two years suggest the second force is weakening. Live revenue has largely stalled, which increasingly looks like market maturation rather than a passing hiccup.

By 2023, most of the addressable audience outside China had already encountered online casino - often through offshore sites with fewer friction points, amplified by aggressive post-COVID marketing. Since only a fraction of people are naturally interested in casino-style gambling, incremental net-new user growth may simply be harder to find. Meanwhile, the remaining “clean” growth pockets are less accessible for a large, regulated supplier without raising compliance and reputational risk, and newer gambling formats (prediction markets, crypto-native schemes) are competing for share.

That leaves a business that can still be best-in-class, but with fewer easy growth paths and a risk profile that matters more when the upside is no longer expanding.

Where Q4 Hits My Q3 Thesis

In Q3, I argued the key question was no longer “will the headwinds fade?” but “is this the new operating baseline?” Q4 pushes the answer toward the latter.

Europe

Ring-fencing is no longer the whole story; channelization is. Management framed Europe’s weakness as regulation pushing players from licensed operators toward unlicensed ones Evolution won’t serve. CEO Martin Carlesund: “players are pushed out of the regulatory remit”, with the addressable market dropping “to 50% level in certain countries”.

That’s the same dynamic I flagged in Q1 - ring-fencing bites hardest where channelization is already low. Q4 makes it look less like a short-term adjustment and more like a structural shrink in the regulated pool.

Asia

Q4 was better than Q3, but not fixed. Management still describes progress as “slow, methodical”, and Carlesund reiterated they “don’t know exactly when that will happen”. In Q1, the message was “great potential” once Evolution “come[s] to terms with […] cyber criminality”; by Q4 the timeline remains open-ended, which is why I no longer underwrite a clean rebound.

It’s also plausible that piracy isn’t the only headwind - maturity may be part of the story. Either way, Asia snapping back to old growth no longer feels like a base case.

Margins / Operating Leverage

Management reiterated margins “in line with 2025 meaning 66%” for 2026. For me, the key is operating leverage.

Profit per table fell to €877k in 2025 (from €1.08 million in 2024). Evolution’s Live model is personnel- and capacity-heavy, so costs are sticky once studios/tables are deployed. In a growth regime, that structure creates strong drop-through; in a flat regime, unit economics compress. The profit-per-table decline is the unit-economics proof.

Management effectively acknowledged this in Q&A: higher volume would have meant “higher volume on the same capacity”. Without renewed Live momentum, operating leverage isn’t working in shareholders’ favor.

The New Disclosure: Customer vs Player IP

The customer-location view is Europe-heavy (71% Europe), but the player-IP view shows Asia and Europe as the two largest demand pools (38% and 35%).

It’s worth reconciling Evolution’s “regulated” language with this disclosure. Evolution only supplies licensed operators, so in a B2B sense “all revenue is regulated”. Yet, the player-IP estimate implies only 47% of end-player activity was in regulated markets in Q4 2025 - highlighting how much underlying demand sits outside well-functioning regulated channels. That is where ring-fencing and declining channelization bite. Evolution’s compliance stance turns market leakage into lost addressable demand.

Share

“New World vs Old World” Framing

Ali Gündüz’s write-up frames Evolution’s reality as bifurcating: “older markets in Asia and Europe face stagnation and even decline”, while the Americas and Africa provide needed growth. The part that matters for me is not just growth rates, but the economic model.

The “Old World” (Europe and Asia) benefited from scale and operating leverage: serving many markets from a few central studios. Meanwhile, major “New World” (North America, LatAm and Africa) markets require local studios (U.S. state-by-state; Brazil language/localization), structurally limiting leverage. Europe and Asia are still the majority of end-demand (73% combined in the player-IP split). The Americas and Africa could grow fast, but from a smaller base they need to grow exceptionally quickly just to offset stagnation in the “Old World”.

Evolution’s own year-end report contains the strategic breadcrumb that aligns with this: “Our primary focus will be on the USA, Latin America”.

Even if the Americas grow, the question becomes whether they can offset “Old World” declines and do so with comparable margins and cash generation. Ali’s conclusion is that this is “difficult to imagine”. I agree.

Profitability and Capital Allocation: Good, But Not Thesis-Saving

Earnings per share (EPS) came in at €5.23 (-11.46% YoY). Net income-to-OCF cash conversion remained solid at 1x, and the balance sheet stayed strong (cash €818 million at year-end) with a debt ratio of 0.26x. Buybacks continued (€500.2 million in 2025) and so did dividends (€572.5 million in 2025).

Strong cash generation is no longer sufficient if the question is durability. In my Q3 article, I said the company was “managing for stability rather than market share gains”. Q4 did not change that characterization.

Ali adds another nuance worth keeping in mind: €200 million of “ghost taxes” from Pillar Two creating a gap between reported tax expense and cash taxes, with uncertainty on timing/settlement.

Why I Exited After Q4

What changed since Q1 isn’t that new risks emerged - it’s that the “temporary” label stopped fitting. I was willing to treat Asia mitigation and Europe ring-fencing as short-term measures that would strengthen the foundation; by year-end, both remained unresolved, and the economics reflected it.

Q4 reinforced the opposite of what the re-acceleration thesis required: Europe is framed as a channelization problem (“regulated markets are losing ground”), Asia remains “slow, methodical” with no timeline, and incremental growth is shifting to lower-leverage regions. Add unresolved distractions (ongoing Playtech litigation, rising headcount, and a weaker H2 Game Round Index due to Europe/Asia measures), and the setup depends too much on external stability.

The most telling line in Q&A was what it would take for operational progress to match financial results: “just a stable environment […] No more, no less”. If re-acceleration depends primarily on external stability rather than controllable execution, I don’t see an asymmetric risk/reward anymore.

Final Thoughts

Evolution remains a quality operator with a strong balance sheet and meaningful capital returns. Q4 reinforced the new baseline: Europe is constrained by channelization decline, Asia has no clear timeline beyond “slow, methodical” progress, and growth is shifting to lower-leverage regions that can’t easily offset weakness in Europe/Asia. If re-acceleration depends mainly on “a stable environment”, the risk/reward is no longer asymmetric. That’s why I exited after Q4 at a -25.8% internal rate of return (IRR).

“We expect to invest about $200 billion in capital expenditures across Amazon, but predominantly in AWS because we have very high demand, […] and we’re monetizing capacity as fast as we can install it.”

— Andy Jassy, Amazon CEO

Amazon’s () Q4 2025 results were fundamentally solid: double-digit revenue growth, accelerating AWS growth, and rising operating income, but the market’s attention snapped to one number. The company expects to invest $200 billion in capital expenditures (CapEx) during 2026.

The market selloff didn’t happen in a vacuum. The market has been in a “CapEx panic” regime: Big Tech has been getting hit whether it spends a lot or a little, and the debate has turned into a referendum on whether hyperscalers can earn attractive returns on an unprecedented scale of investment.

Amazon was no exception. In after-hours trading immediately after the earnings release (Feb 5, 2026), shares fell about 11% as investors reacted to the magnitude of the 2026 CapEx plan - up 56% from 2025.

The near-term story isn’t “demand is weak”. It’s that Amazon is choosing to convert today’s cash-generating power into capacity (AI infrastructure, chips, robotics, and low Earth orbit satellites), and investors are repricing the uncertainty around how quickly that capacity translates into incremental profit.

In this article, I analyze Amazon’s financial performance, its business segments, and why I remain bullish despite near-term volatility, while also laying out the signposts that will determine whether the $200 billion posture is disciplined or reckless.

Quarterly Highlights

The Q4 2025 results had plenty of moving parts, but four data points captured my attention:

1. Revenue stayed in double-digit growth territory: sales grew to $213.4 billion, up 14% year over year (YoY). Beyond segment reporting, the internal mix continued shifting toward services: Q4’s advertising services rose to $21.3 billion (+23% YoY) and third-party seller services rose to $52.8 billion (+11% YoY).

2. Operating income rose to $25 billion, up 18% YoY, but Q4 included notable special charges - $2.4 billion in tax dispute resolution/litigation settlement, severance, and physical-store impairments. Stripping those out, operating income would have been $27.4 (+29.2% YoY), useful context when comparing quarter-to-quarter margin narratives.

3. The engine that justifies the investment cycle accelerated: AWS revenue grew 24% YoY to $35.6 billion, its fastest growth in over 3 years. In Jassy own words, “AWS is now a $142 billion annualized run rate business”. AWS operating margin was 35%, down from 36.9% in the prior-year quarter. Margin matters because a meaningful portion of the $200 billion build-out will flow through AWS depreciation and operating expense, and the market is already testing whether AI-era cloud returns can stay attractive.

4. AWS is larger than Azure and Google Cloud. So even if peers post higher percentage growth in a given quarter, AWS is still adding massive absolute dollars. Jassy noted that “[…] it’s very different having 24% year-over-year growth on a $142 billion annualized run rate than to have a higher percentage growth on a meaningfully smaller base”. He emphasized AWS’s embeddedness in large enterprises, noting that more S&P 500 companies run “AWS as their primary cloud provider” than the next two largest competitors combined - a moat that’s hard to ignore when evaluating whether demand is real or speculative.

Share

Business Segments’ Performance in 2025

The full-year 2025 results clarify the core reality of Amazon’s model: retail is the scale flywheel with improving profitability, AWS is the operating profit engine, and services (ads, third-party, subscriptions) increasingly subsidize both.

At the consolidated level, 2025 revenue was $716.9 billion (+12% YoY) and operating cash flow (OCF) was $139.5 billion (+20% YoY), a 19.5% OCF margin, while free cash flow (FCF) fell to $11.2 billion. Amazon explicitly attributed the FCF decline to a $50.7 billion YoY increase in CapEx, stating that the increase “primarily reflects investments in artificial intelligence”.

By segment for the year:

• North America delivered $426.3 billion of revenue (+10% YoY) with operating income of $29.6 billion (+18.4% YoY), a 6.9% operating margin. This is the “quiet win”: Amazon keeps extracting efficiency from a logistics machine that many still think of as structurally low-margin.

• International produced $161.9 billion of revenue (+13% YoY) with operating income of $4.7 billion (+23.7% YoY), a 2.9% operating margin. International remains the least “finished” part of the story - progress is real, but the segment still absorbs local fulfillment build-out and is more exposed to local execution and macro variability.

• AWS delivered $128.7 billion of revenue (+20% YoY) with operating income of $45.6 billion (+14.6% YoY), a 35.4% operating margin. AWS generated about 57% of total operating income while representing about 18% of total revenue. This is what gives Amazon the confidence to press the spending pedal harder and tolerate near-term FCF compression to protect and extend AWS leadership, even as the market demands proof that margins won’t structurally reset lower.

Beyond the segment numbers, services are increasingly the margin layer. Advertising grew 23% YoY in Q4 to roughly $21.3 billion, helped by Prime Video’s ad tiers, and Amazon is reducing advertiser friction by unifying Demand Side Platform (DSP) and Sponsored Ads into a single campaign workflow with AI-assisted tools.

Finally, Amazon is pushing AI monetization inside retail, not just AWS. Jassy emphasized:

The CEO also highlighted agentic shopping features like Rufus, noting users are more likely to complete a purchase and can automate tasks like research, price tracking, and auto-buy, potentially turning AI spend into higher conversion and more high-margin attach over time.

Capital Intensity and Balance Sheet

Capital intensity is visible on the balance sheet.

Gross property and equipment rose to $534.1 billion (+35.5% YoY) in 2025, while net property and equipment increased to $357 billion (+41.3% YoY). Within that, servers and networking equipment increased to $172.5 billion (+52.4% YoY) and construction-in-progress increased to $71.7 billion (+53.9% YoY) - two line items that are consistent with hyperscaler data center build-outs, and the supporting “assembly line” behind them: land, concrete, power, cooling, networking, switches, memory, chips, and GPUs.

Depreciation and amortization expense on property and equipment rose to $41.9 billion (+30.5% YoY) in 2025, which provides a preview of the net income margin headwind that will grow if CapEx continues to accelerate. This is one reason some investors get uncomfortable: even if the cash is there, the accounting drag becomes more visible before the payoff does.

The balance sheet shows funding flexibility, even in a heavy-investment year. At year-end, Amazon reported $123 billion of cash, cash equivalents, and marketable securities (net of restricted amounts).

Long-term debt was $65.6 billion, and Amazon disclosed it had $68 billion of unsecured senior notes outstanding, including $15 billion issued in November 2025 for general corporate purposes.

The company also reported no share buybacks, which matters because it keeps “dry powder” available for CapEx at the expense of near-term capital returns. That choice is part of the Amazon identity, and one reason Amazon can outspend peers without immediately turning the conversation into dividends, buybacks, and capital return tradeoffs.

Share Level-Headed Investing

Interpreting the 2026 CapEx Plan

Amazon’s $200 billion 2026 CapEx guide is mostly a message-versus-optics issue. Management framed the step-up as demand-driven, not speculative: Jassy anticipates “strong long-term return on invested capital” and he emphasized the spend will be “predominantly in AWS”, because customers “really want AWS for core and AI workloads” and Amazon is “monetizing capacity as fast as we can install it”.

The selloff reflects investors’ sentiment, not a solvency issue. $200 billion is well above expectations and, by definition, compresses near-term FCF, fueling the broader “Big Tech CapEx crowds out short-term returns” narrative. The better way to view it is against Amazon’s run-rate: cash CapEx rose from $77.7 billion (2024) to $128.3 billion (2025), and $200 billion implies another 56% step-up. Amazon can fund that with strong OCF and a liquid balance sheet; the uncertainty is payback timing as depreciation rises.

The bull case is demand visibility plus better unit economics. Amazon disclosed in the 2025 10-K filing $244 billion of long-duration performance obligations (mostly AWS) with a 4.1-year life, a backlog-like support for a multi-year build if utilization and pricing hold.

Jassy also framed demand as “barbelled”: AI labs as near-term anchor tenants, with enterprise productivity/cost-avoidance workloads broadening over time. On economics, Trainium + Graviton are now above $10 billion run-rate and growing triple digits; Trainium2 is sold out (1.4 million chips) and underpins most Bedrock inference (100,000+ customers), with Trainium3 largely committed by mid-2026 and Trainium4 coming in 2027. Graviton is up to 40% more price-performant than leading x86 chips, helping bend AWS’s cost curve.

The bear case is classic overbuild risk in a crowded cycle: power/chip constraints and price pressure can compress returns before they show up. If OCF stays near $140 billion and cash CapEx hits $200 billion, FCF likely goes negative, shifting scrutiny to funding mix (cash vs. debt) and whether ROIC holds.

Key Signposts to Watch Through 2026

If the “catalysts” case is the strategy for years to come, 2026 is about widening the business’ moats. A $200 billion CapEx posture is disciplined only if incremental capacity converts into durable AWS monetization and acceptable incremental returns once depreciation and ramp costs hit the P&L.

First, AWS growth and margin durability. Amazon is tying most spend to AWS and saying demand is the constraint. The signpost is sustained re-acceleration as capacity comes online, without a lasting margin reset downward after the build-out peak.

Second, backlog quality. The $244 billion of long-duration performance obligations is the closest thing to a backlog signal. I will be watching whether it keeps expanding and whether the remaining life stays healthy - those two metrics shape confidence in utilization and pricing.

Third, custom silicon commercialization. Trainium and Graviton matter most if they scale as products, not just internal advantage. The signpost is adoption into major workloads and growing revenue mix, because that’s how Amazon bends the compute cost curve in its favor.

Fourth, cash versus optics. I expect 2026 margin noise as depreciation and ramp costs rise. I don’t anchor on FCF in isolation; I anchor on OCF and the balance sheet path, especially whether any FCF gap is funded by cash drawdown or incremental debt.

Valuation: Is Amazon Undervalued?

With CapEx spiking, I value Amazon on OCF, not on GAAP earnings or near-term FCF. OCF is cleaner, and it’s still compounding. In 2025, Amazon generated $139.5 billion of OCF (+20% YoY). The chart also shows why the stock reacted so violently: while OCF has marched higher, the P/OCF multiple has compressed sharply. At 15.3x today, it’s the low end of the past decade range and well below the historical median of 25x.

To translate that into an underwriting-style fair value, here’s a simple OCF framework:

• Start with 2025 OCF of $139.5 billion and grow it at 20% per year for five years: OCF reaches $347.2 billion in year five.

• Assume dilution of 1.15% per year: shares rise from 10.863 billion to 11.5 billion.

• Apply a 20x P/OCF multiple (below the longer-run median): implied year-five price is $603.64.

• Discount back at 10% for five years: present fair value is $374.81 per share.

This isn’t a target price; it’s a sanity check on what the stock could be worth if OCF continues compounding at a strong rate and the market rerates the multiple closer to a more normal band once the CapEx cycle is better understood. The risk is straightforward: if the $200 billion build leads to a structural ROIC or margin reset in AWS (or OCF growth slows materially), the multiple should stay depressed, and the math changes quickly.

Despite short-term market concerns, my long-term investment thesis remains intact. Should negative sentiment persist, I will consider increasing my position, even though it is already my second-largest holding.

Final Thoughts

The market’s reaction to the earnings report wasn’t really about the quarter. It was about underwriting a steeper investment slope. A $200 billion 2026 CapEx guide forces investors to choose between two narratives: “Amazon is overbuilding into a crowded AI cycle”, or “Amazon is pulling forward capacity into demand it can already see”.

Amazon’s Q4 2025 didn’t read like a business losing momentum. It read like a business choosing to compress near-term FCF to extend its lead in infrastructure and AI-adjacent services, and asking shareholders to underwrite the timing risk. Whether that’s attractive depends on your tolerance for a “build now, harvest later” cycle, but the underlying operating trajectory (especially AWS re-acceleration and OCF growth) is the part I treat as signal.

I still lean toward the disciplined-investment interpretation. The operating picture is solid, and the balance sheet gives Amazon room to invest without turning this into a financing story. If the incremental spend shows up as sustained AWS growth with resilient margins, expanding long-term commitments, and rising commercial adoption of Trainium/Graviton, the 2026 CapEx plan will look less like a bet and more like a deliberate move to extend cost and scale advantages.

Investors are right to demand proof. Depreciation and ramp costs will pressure reported margins, and FCF may look worse before it looks better. That’s why the signposts matter. Through 2026, the debate won’t be settled by one quarter’s FCF print; it will be settled by whether Amazon converts capacity into durable earnings power, without sacrificing pricing discipline or letting the build-out reset AWS returns structurally lower.

“Invert, always invert.”
— Charlie Munger

In two articles before this one, I tried to apply that idea in a practical way: instead of grabbing the most convenient narrative, I deliberately argued both sides of the same AI CapEx debate and forced myself to keep them consistent.

In Are AI Chip “Useful Lives” Creating Useless Earnings?, I took Michael Burry’s claim seriously: hyperscalers may be flattering reported earnings by stretching server/GPU useful lives just as AI CapEx explodes. Even if that’s not “cooking the books”, it forces a level-headed investor to separate earnings optics from economic reality, and to anchor on cash (OCF and FCF), not EPS.

In Laying the Tracks of the Digital Economy, I argued the opposite-sounding point: the AI and cloud build-out is a real, physical infrastructure race - more like railroads than software - and the upfront capital expenditure (CapEx) is the price of securing long-term dominance. The tracks are fiber and power; the locomotives are AI training and inference; the toll booths are cloud platforms.

Pershing Square’s messaging during their February 11, 2026 annual investor presentation is useful because it stitches these two threads together cleanly. Their thesis is not “CapEx is good”. It’s: for the right businesses, today’s AI CapEx is not reckless spending or accounting theater - it’s demand-driven capacity expansion that widens moats. And that distinction matters.

The market is treating today’s AI CapEx like unavoidable maintenance spend, when for the dominant hyperscalers it is largely growth CapEx pulled forward by real demand. If that’s correct, higher CapEx today should translate into greater capacity, higher revenue, and stronger long-term cash generation - especially for the platforms with distribution and data. But the proof will not show up in manicured EPS. It will show up in the long-run relationship between OCF, CapEx, and (eventually) FCF.

The Market’s Mistake: Treating Growth CapEx Like Maintenance CapEx

Pershing’s core point is simple and it maps directly to the tension I highlighted in the Burry article.

If maintenance CapEx is rising (money you must spend just to keep the business at its current level) owner earnings are structurally worse. That’s a real hit to intrinsic value. If growth CapEx is rising because demand is exceeding capacity, and incremental capacity can be monetized at attractive returns, then the spending is a feature, not a bug. In that case, the investor’s reaction is closer to what I described in the railroad analogy: you don’t complain that a rail company is laying track during a land grab; you ask whether it’s laying track on the right routes, with full trains likely to follow.

Pershing Square founder Bill Ackman said this distinction bluntly:

But when the spend is driven by demand, the reaction should invert:

Their punchline is exactly the one the market often misses:

This is where many “AI CapEx freakouts” go wrong. They price the spending as a perpetual sustaining burden. Pershing frames it as a bottleneck relief investment: build capacity now because customers are already pulling it through.

Supply Constrained Is the Tell

In my infrastructure piece, the recurring theme across AWS, Azure, and GCP was capacity constraints. Amazon’s CEO Andy Jassy said AWS “could be growing faster” if not for capacity constraints. Microsoft’s CFO Amy Hood said AI demand was “higher than our available capacity”. Google’s CFO Anat Ashkenazi said Google “exited the year with more demand than we had available capacity”. The same was true in the Q4 2025 earnings calls.

Pershing treats “supply constrained” as the tell, and they say it explicitly. On AWS, Pershing’s CIO Ryan Israel described it in plain language: “Demand was much higher than supply. They weren’t able to meet that demand” and that’s why Amazon is investing heavily. More broadly, he said: “What a lot of these hyperscalers are spending money on is physically trying to build more infrastructure so they can serve all the demand that they need today”.

Pershing’s view is basically: if you believe those statements are directionally true, then elevated CapEx is not primarily a margin problem. It’s an unmet demand problem.

This is also why Pershing’s take complements (and doesn’t contradict) the Burry piece. Burry is warning that utilization does not equal value creation - busy GPUs can still be economically mediocre once you include power, cooling, and opportunity cost. Pershing is arguing that the strongest evidence against “empty trains” is the opposite: customers are waiting on capacity. That doesn’t prove returns will be excellent, but it shifts the base rate. If demand is genuinely pulling, the probability of decent incremental returns rises.

Where Burry stays relevant is in forcing the next question: if demand is there, what are the returns on the incremental dollar, net of the full cash outlay?

Share Level-Headed Investing

A Mini-Case: What “Proof” Would Look Like in Practice

Take AWS, because it cleanly fits Pershing’s framing. Israel’s point was that demand was already above supply and Amazon “wasn’t able to meet that demand”. If that’s true, the test isn’t philosophical. It’s operational:

• First, capacity additions should show up as a re-acceleration (or at least resilience) in cloud growth versus what you’d expect without the extra infrastructure. AWS’s growth has looked more resilient recently than the market seemed to expect.

• Second, incremental margins should not collapse. If a company adds capacity into genuine backlog, the new dollars should be relatively high quality. If margins crater, it’s a sign the spend is fighting commoditization or misallocated build-out.

• Third, over a longer arc, OCF should rise faster than CapEx as the build phase matures. In a railroad boom, the build is front-loaded. In a treadmill, the “build phase” never ends.

In other words: if the bottleneck is real, removing it should be visible. If it isn’t visible, the market’s “maintenance CapEx” suspicion starts looking more reasonable.

Why Pershing Focuses On the Mega-Caps

Pershing isn’t making a generic “AI will be big” bet. They’re making a competitive structure call: in an era where AI is capital-intensive and physically constrained, the advantage shifts toward the platforms with distribution, proprietary data, and balance-sheet firepower. They’re explicit that this is not true for every company. Israel: “I wouldn’t say that every company that’s going to spend a lot of money on AI I would be supportive of”. The “right businesses” are the incumbents with moats already in place:

That lines up with the thesis underneath both of my earlier articles:

1. Scale is not optional anymore: In a railroad-style build-out, small players don’t innovate their way around steel, land, and capital. In AI infrastructure, you don’t iterate your way around power, GPUs/ASICs, networking, and data centers. Pershing’s point is that “the larger, more differentiated businesses have the opportunity to put in more capital”, and that this is self-reinforcing because “that is going to further allow them to embed themselves with their customers”. The hyperscalers can fund the build internally; many adjacent players cannot.

2. Distribution converts CapEx into cash faster: A dollar of AI capacity is worth more when it can be deployed into an existing engine - billions of users (Search, YouTube, Meta’s “Family of Apps”), sticky enterprise workflows (Microsoft), or a dominant cloud utility (AWS, GCP, Azure). Pershing’s reason is simple: the leaders already “have the distribution […] and the data”, so new capacity gets absorbed faster and turns into customer lock-in. The same spend at a narrower business can be value-destructive.

3. Data and feedback loops make AI less commoditized: In the railroad analogy, owning the right routes mattered. In AI, owning the right data and feedback loops matters. Better inference, better personalization, better ad performance, better conversion - those are monetization rails. Pershing’s checklist is essentially the same: proprietary data, being deeply embedded, and distribution, because those moats make AI “incredibly enhancing” rather than commoditizing.

This is the bridge between my two articles: even if AI creates “useless earnings” in parts of the ecosystem, it can widen moats at the aggregators because those platforms can translate capacity into durable monetization.

Subscribe now

A Level-Headed Framework: Reconciling Pershing with Burry

This is mainly a hyperscaler/aggregator framework. It applies to businesses that already have massive distribution, proprietary data at scale, and the balance sheet to self-fund the build-out. It does not automatically apply to every “AI beneficiary”, especially downstream companies that must buy compute at market prices, compete on features alone, or rely on external capital. In a world where power, chips, and data centers are binding constraints, second-derivative players can still win, but their economics are far more fragile.

Pershing Square is directionally right if:

• The company is genuinely supply constrained (not “PR constrained”).

• There are clear demand signals: backlog, contracted commitments, utilization, customer pipeline.

• Incremental capacity is monetized through an existing cash engine (cloud contracts, ads, commerce).

• The company can self-fund the cycle without financial fragility.

Burry is directionally right if:

• Growth CapEx quietly becomes sustaining CapEx, with no plateau.

• Useful lives are extended beyond economic reality, flattering near-term operating income.

• The cash gap widens: OCF grows, but CapEx grows faster for longer, keeping FCF thin.

• Impairments/accelerated depreciation start showing up (the “stranded assets” risk).

In my Burry article, I pointed out the exact place the truth will surface: not in manicured EPS, but in the long-run relationship between operating cash flow growth and CapEx. Pershing’s thesis doesn’t change that. It tells me what to expect if the bull case is right: “CapEx now, FCF later” should eventually become visible in the numbers.

They collide first in depreciation/EPS optics, and then get settled in the cash flow statement - OCF versus CapEx, and eventually FCF.

How the Bull Case Breaks

Even if demand is real today, this cycle can still disappoint investors. Here are the two failure modes that matter most:

1. Demand pullback through efficiency: Models get cheaper to run, inference becomes more efficient, customers optimize usage, and the same work gets done with fewer compute dollars. In that world, yesterday’s backlog doesn’t automatically justify tomorrow’s capacity.

2. Pricing pressure and commoditization: If hyperscalers and competitors add capacity faster than demand grows, unit pricing can fall. A company can be busy and still destroy value if the marginal dollar is competed away.

Those risks tell me what to watch: not “Is AI big?” but “Do the unit economics and cash returns survive scale?”

What would change my mind is not a new narrative, but a clean mismatch between capacity talk and the numbers. If hyperscalers stay “supply constrained” while growth keeps decelerating as new capacity comes online, I’ll treat the spend as closer to sustaining CapEx (or misallocated growth CapEx). If, instead, capacity additions are followed by resilient growth, stable unit economics, and a visible shift where OCF begins to outpace CapEx as the build matures, then the market’s “maintenance CapEx” framing will have been the mistake.

Final Thoughts

My railroad-analogy article made the case that the hyperscalers are laying the tracks of the digital economy, and that the scale of CapEx is itself a barrier to entry. My useful lives article warned that the same CapEx cycle can flatter earnings and hide the true economic burden, making cash the only trustworthy anchor.

Pershing Square’s contribution is the missing framing that links them, and it’s best captured in their own words. If the spending is maintenance, “the stock should go down a lot”. If it’s demand-driven capacity expansion, “you should applaud when they’re doubling the amount of money that they’re spending”. Their bet is that, for the dominant platforms, today’s AI CapEx is “squarely in the growth CapEx category”, because “demand was much higher than supply” and the leaders are racing to remove a real bottleneck.

The discipline remains the same. The market will argue about narratives; the cash flow statement will settle it.

If AI CapEx is truly growth CapEx, we should eventually see it in accelerating operating cash flow, stabilizing CapEx growth, and widening long-term free cash flow. If it’s really sustaining CapEx with optimistic useful lives, the cracks will show up in persistent FCF compression, shorter depreciation lives, impairments, or both.

Level-headed investing here is not choosing between Burry and Pershing. It’s using Burry’s skepticism to avoid being fooled by optics, and using Pershing’s framework to understand why the best “rails owners” may emerge from the CapEx mega-cycle stronger than before.

Over the past year, fears that AI will commoditize software triggered indiscriminate selling, as if “software” were a single risk bucket. That misses what CSU actually owns. The irony is that this has been one of the sharpest sentiment drawdowns the stock has experienced, even as the underlying machine has kept compounding. When price and fundamentals diverge this hard, it’s usually not because the business suddenly broke. It’s because narratives moved faster than cash flows.

AI may lower the cost of writing code, but it doesn’t erase operational risk, switching costs, or the decades of workflow knowledge embedded in vertical market software (VMS). The thesis breaks if deal discipline erodes at scale, or if AI turns costs variable faster than CSU can price and route around it.

A Decentralized Reinvestment Flywheel

Most companies face a reinvestment problem: as they scale, high-return opportunities shrink and surplus cash gets misallocated on vanity projects, empire-building, or buybacks done at elevated share prices. Constellation solved that by industrializing reinvestment inside a decentralized structure. The model is simple:

1. Acquire mission-critical vertical market software (VMS) businesses.

2. Harvest their steady cash flows.

3. Reinvest those cash flows into more acquisitions.

4. Repeat.

That loop, done with discipline, has produced one of the most exceptional long-term records in public markets: since CSU’s 2006 IPO, it has compounded at 29.3% annually, turning a $10,000 investment into more than $1,600,000 in less than 20 years. CSU has become a textbook example of the serial acquirer model.

The past matters here not because I’m extrapolating a straight line, but because process is the product. CSU’s playbook has been tested across cycles and thousands of acquisition decisions: buy resilient, sticky, cash-generative businesses; stay disciplined on price; avoid “integration” that destroys what made the asset valuable; and fund the next cycle largely from internal cash flow rather than external capital. Because CSU is a perpetual owner, founders can sell without feeling like they’re handing their life’s work to a financial engineer planning to flip it in three years.

There’s another layer that matters just as much as discipline: sourcing. At the holding company level, the business looks easy to copy in theory. Anyone with a checkbook can buy small software companies. In practice, the hard part is distribution in M&A: finding thousands of niche targets, building credibility, and staying top-of-mind for years until a founder is ready to sell. CSU’s edge isn’t that competitors can’t write checks. It’s that CSU has built an acquisition pipeline, a reputation for stewardship, and an operating model that sellers trust.

On September 25, 2025 succession headlines, the market panicked when Mark Leonard stepped down as President and Mark Miller was appointed. Leonard shaped the culture, but CSU was built to outlast any one person. Authority is cascaded down across operating groups and business units. The company’s real intellectual property isn’t the code-base; it’s a culture of aligned incentives, acquisition discipline, and capital allocation rigor. That human-capital moat doesn’t show up neatly on a balance sheet, but it’s the reason the system scales at all.

The Fortress Economics: VMS is Not “Software” in The Generic Sense

To understand why Constellation compounds1, start with what it buys.

Horizontal software can be excellent, but it often sits closer to the surface: generalized tools for broad workflows like CRMs, productivity suites, and collaboration tools, with more direct competition and lower switching friction. Vertical market software (VMS) is different. It’s built for a specific industry - think municipal billing, utilities, dispatch systems, library systems, niche healthcare workflows, other unglamorous back-office or operational systems, with edge cases that only appear after years in production. It’s the difference between software that helps you operate and software you operate through.

The total addressable market is smaller, but the product is embedded and the switching costs are higher. Customers are conservative because failure isn’t an inconvenience; it’s downtime, compliance risk, and career risk. In many verticals, these systems are customized, integrated, implemented, and supported over years. In plenty of cases the “product” includes professional services, onboarding, and ongoing fixes. In some corners, legacy on-premise deployments and entrenched workflows make replacement even more painful. The moat is as much implementation and accountability as it is source code.

These niches also have a structural defense: many are too small to justify venture capital attention. If the prize is modest, rational competitors don’t spend years burning cash for a limited payoff. That’s how CSU ends up owning hundreds of “tiny monopolies” in boring corners of the economy - boring in narrative, excellent in economics.

Many of these businesses are not organic growth machines. Organic growth can be modest because the market is mature and the vendor already dominates it. Pricing may rise gradually and churn is often driven by customer failures rather than competitive losses. The genius is not “these assets grow fast”. The genius is “these assets produce cash reliably”, and CSU has an engine to redeploy that cash into the next purchase.

What Breaks The Thesis

Owning CSU responsibly means being explicit about what can change.

1. Deal discipline at scale. Scale is a constraint. As CSU grows, it must find larger deals, broaden the opportunity set, or accept lower incremental returns. During an investor update call in mid-September 2025, management has been candid about being “opportunity constrained” and they’ve been equally clear that sitting on cash is not the plan. The risk is straightforward: scale pressure slowly pushes the organization into higher prices, larger checks, or categories with weaker economics, and the return profile of the loop degrades over time. This is also why deal size matters. The historic “bread and butter” has been many small acquisitions. But as free cash flow grows, funding the same pace with only tiny deals becomes harder. One way to extend the runway is occasional larger acquisitions (sometimes supported by measured leverage). That can be rational if hurdle rates hold. It’s dangerous if larger deal size becomes a justification for lower standards.

2. Low-switching-cost pockets and “graduation risk”. Not every niche is equally mission-critical. CSU has long acknowledged that large, sophisticated customers can leave, sometimes to SAP implementations or proprietary builds. AI increases the stakes because it lowers the perceived friction of “build vs. buy” and can shift the bargaining line over who does customization work. The risk concentrates at the top end and in less critical niches, where switching costs are genuinely lower and buyer sophistication is higher. This is portfolio risk, not an existential one, but it matters if it spreads. Leonard also flagged that some categories are more exposed than others - customer service software, for example, is already seeing more pressure - so the key is not pretending every corner of CSU’s portfolio has identical switching costs.

3. Variable-cost pressure and centralization creep: AI introduces a new cost structure. Software has historically had high incremental margins; AI can add a usage-based COGS element (model access, compute, inference). Today’s AI usage is heavily subsidized by model providers competing for market share. Over time, those providers may try to recapture that investment through pricing, and vendors like CSU may find a larger slice of their value captured upstream. CSU can route around this, through pricing, model selection, selective on-premise inference, and disciplined productization, but it’s a real economic variable. Separately, AI may require shared infrastructure (security guardrails, common platforms). Some centralization is rational. The execution risk is cultural: if centralization spreads beyond what’s necessary, it can dull the autonomy and accountability that make the system work.

The Trust Moat vs. The AI Scare

The loudest bear case is that AI will let customers code their own software, making CSU’s portfolio obsolete. It sounds plausible until you think about incentives and accountability.

Reliability beats sexiness. Many CSU end-markets are places where probabilistic failure is unacceptable. Utilities, municipal systems, and operational workflows don’t get swapped lightly for “AI-first” tools that might mishandle edge cases. In these niches, trust is the product.

The economics don’t reward switching. For many customers, software is a small line item relative to revenue, often under 1%. Cutting the bill in half often moves margins by basis points, while the downside is asymmetric: downtime, compliance/security exposure, and reputational risk for the buyer. CSU isn’t seeing customers “vibe code” their way out of core systems at scale; most want to run their businesses, not own software maintenance for the next decade.

AI makes code easier to produce, not consequences easier to own. The real question isn’t “can AI write code?” It’s “who is accountable when it breaks?” What’s sticky in VMS isn’t just customer data. It’s decades of accumulated workflow knowledge: business rules, exceptions, integrations, and the lived reality of operations, embedded into systems customers rely on daily. A newcomer can write a cleaner UI. That’s not the same as installing, integrating, servicing, and standing behind a system that runs payroll, billing, dispatch, or compliance.

This is why many VMS businesses settle into tiny monopolies or duopolies with low churn. The threat isn’t that AI can write code. The threat is whether AI can erase switching costs and operational risk. That’s a much higher bar.

How CSU Can Use AI Without Betting The Company

The sensible base case is augmentation. Leonard framed it with intellectual honesty the call by reminding investors how often “replacement” forecasts miss: “Geoff [Hinton] wasn’t wrong […] Where he was wrong was that the technology would replace people. Instead, it’s augmented people”. Applied to software, the honest answer is uncertainty: AI could raise throughput (getting more done with the same teams) dramatically, but it can also raise long-run maintenance costs if it encourages more code, more complexity, and less coherence.

That’s exactly the kind of second-order thinking I want from a perpetual owner of mission-critical software.

CSU’s operating posture fits its culture. It’s not a centralized moonshot. It’s hundreds of business units testing AI where it’s useful - testing, documentation, internal tooling, selective refactoring - while keeping a tight leash on customer-facing reliability. Management has pointed to concrete areas like better unit testing and security improvements (including vulnerability detection), while acknowledging real constraints: huge legacy code-bases can exceed today’s practical context windows, and “rewrite everything” still leaves you with a system you must maintain.

The adoption path stays conservative because CSU’s customers are conservative. Many prefer their vendor to embed AI into systems that already work, rather than rip-and-replace core infrastructure. The practical playbook is simple and unglamorous: spend time with customers, find high-ROI pain points, and ship changes that reduce regret.

The Opportunity in The Noise: AI Fear Can Feed The Flywheel

One thing I respect about CSU is its temperament. There’s enormous incentive right now for CEOs to sell a sweeping AI story. Reality is messier. CSU’s willingness to say “we don’t know yet” may be bad for a one-day stock move, but it’s the right stance for a perpetual owner trying to compound for decades.

The market is missing a key point: Constellation is a buyer of small software businesses, not a seller of hype. When fear rises - AI fear, succession headlines, generalized “software is dead” narratives - valuations in the small-software ecosystem can soften. That isn’t automatically bad for CSU. It can be fuel. If anxious sellers accept lower multiples, CSU gets to deploy capital at better prospective returns. That aligns with the system: high hurdle rates, low tolerance for overpaying, and a long runway of reinvestment. While the market debates software’s terminal value, CSU can keep buying durable cash flows at prices that improve the math of compounding.

Final Thoughts: Don’t Stop the Snowball

Constellation is a filter for the portfolio. It filters out the temptation to chase whatever is loud this quarter and forces attention back to what actually compounds: mission-critical systems, entrenched workflows, and cash flows reinvested with discipline.

At an 8% weighting, my conviction isn’t that CSU will never change. It will. My conviction is that the system is built to adapt without breaking: resilient niche businesses throwing off cash, that cash funding more acquisitions, and a culture designed to preserve what makes the acquired businesses valuable.

If the market wants to treat software as one undifferentiated AI loser, fine. CSU lives in the part of software where trust is the product and switching is the risk. When that gets mispriced, volatility isn’t a signal to panic. It’s the market offering a better entry point into a compounding machine.

Culture is the backdrop to this series. The Lifestyle Inflation Trap mapped the mechanism (upgrades → fixed bills → slack shrinks). The Demographics of Slack and Saving showed who even has margin to protect. And Why Saving Is Hard explained the wiring: why immediate rewards, status cues, and emotional relief make spending feel vivid while saving stays abstract. This piece adds the missing layer: the cultural engine underneath it all.

Western economies run on a kind of hedonic treadmill at the societal level. New products and trends constantly reframe “good enough” as “outdated”: last year’s phone works perfectly well, but the norm moves on; last season’s clothes still fit, but fashion cycles redefine what feels current. That churn isn’t accidental. It keeps demand alive and businesses profitable - because in many economies, consumption is a major driver of GDP. In the US, consumer spending is over two-thirds of GDP, so aggregate growth is heavily tied to households continuing to buy.

When consumption is structurally central to economic performance, you should expect a constant cultural tailwind behind “buy now”.

A simple framing I use: cultural slack is the distance between what your environment pushes you to spend, and what you choose to spend. When cultural slack is low, you need systems and boundaries.

The Western Consumption Machine: “More, Newer, Better”

Post-WWII America didn’t just get richer, it built an identity around mass consumption. In the 1950s, consumer spending was explicitly framed as civic participation: buying was cast as “good citizenship” in a mass-consumption economy. This is one reason the US is often the clearest expression of modern consumer culture: material upgrading became socially legible proof that you were “doing well”, and the suburban/home/car lifestyle became a central script.

This script spread beyond the US. As consumer society matured across industrialized economies, it was supported not only by rising incomes but also by public policy choices and resource abundance that made consumerist lifestyles scalable.

Advertising is not a sideshow in that system. Total advertising has been estimated at 1.3%-2% of US GDP over long periods, which is another way of saying: a meaningful share of the economy is dedicated to influencing attention, preference, and purchase.

If the default cultural message is “upgrade” you should assume your baseline will drift upward unless you build counter-defaults.

Easy Credit: Turning Wants Into Obligations

Consumer culture becomes far more powerful when it’s lubricated by credit. Modern card systems and general-purpose charge cards expanded in the second half of the 20th century; Diners Club (1950) is commonly cited as the first modern general-purpose charge card, and card-based payment systems then scaled quickly.

Credit matters culturally because it breaks the natural boundary between desire and affordability. Credit cards, installment plans, device financing, and BNPL lower the felt cost at the moment of purchase, so you experience the benefit now and push the pain into the future. That’s not just a personal weakness; it’s a product design feature in many consumer markets.

Culture supplies the urge; credit supplies the mechanism that turns urges into defaults. When financing is normal, you don’t need slack to upgrade; you need a monthly payment. That’s how lifestyle creep becomes fixed-cost creep: not one splurge, but the contractualization of your preferences into recurring obligations.

Credit doesn’t just increase consumption; it increases pre-commitment. Future slack gets sold forward.

Subscribe now

Social Media: The Comparison Engine Goes Global

Social comparison is old. Social media industrialized it.

Western consumer culture existed long before Facebook, Instagram and TikTok - what changed is visibility. Social platforms compress comparison into daily life and shift spending from a mostly private choice to a public signal. The behavioral channel is well studied: people compare upward, feel the gap, and then try to close it; sometimes with purchases.

A consistent finding in recent research is that exposure to influencer content and curated consumption increases social comparison and “fear of missing out” (FOMO), and that these mechanisms can predict conspicuous consumption. More broadly, research on social media use links platforms to intensified social comparison and envy dynamics.

This matters because it shifts “normal” from your local peers to an algorithmically curated reference group. If your feed is full of upgraded kitchens, business-class flights, and luxury meals, the baseline quietly moves - even if your income doesn’t.

In other words, social media compresses cultural slack by raising perceived norms faster than real income grows.

Italy: “La Dolce Vita” Meets a Money-Taboo Hangover

Italy is a hybrid case: a cultural bias toward pleasure and presentation (food, social life, style), alongside an older tradition of prudence shaped by scarcity and post-war memory. You can see the tension in real life: “enjoy life” spending coexists with a quiet reluctance to be overtly ostentatious, depending on region, cohort, and context.

Where it gets financially tricky is that while the culture is openly pro-pleasure, family norms are less open about explicit planning. This is reflected in the fact that 40% of Italians aged 18-34 never discuss money at home and feel uncomfortable talking about finances. This discomfort is often linked to a traditional Catholic/Latin context where money talk can be seen as taboo or “greedy”, a sentiment echoed by Giovanna Paladino (Director of the Italian Museum of Saving), who argues that money is frequently viewed with suspicion even though it is fundamentally a tool for achieving life goals.

When consumer culture rises but money skills and money conversation lag, slack leaks faster, because the environment sells upgrades, while the culture doesn’t reliably teach planning.

The UK: Cautious Identity, Credit Reality

The UK sits between American-style consumer finance and a legacy of European caution. There’s still a cultural memory of restraint, but modern households face a high-cost baseline (especially housing) and frequent “just finance it” offers. One way to see the earlier “leverage era” is household debt: UK household debt as a share of income rose sharply from the late 1990s and peaked around 2008.

That history helps explain why the national conversation has shifted toward resilience, savings buffers, and vulnerability under cost-of-living shocks. The FCA’s Financial Lives work repeatedly highlights low emergency buffers for large segments of the population (for example, around one in ten adults with no cash savings, and many more with limited emergency reserves).

Share Level-Headed Investing

Guardrails That Work In a Consumer Culture

I assume the environment is optimized to push spending and normalize upgrades, so I rely on guardrails that reduce exposure, set boundaries early, and keep my baseline from creeping up by default. These are the practices I use to protect slack and preserve financial flexibility in a world that’s constantly asking me to spend a little more:

1. Be explicit and refuse the scoreboard: I’m explicit about what I’m willing to spend on because it genuinely improves my life (traveling, experiences with people I care about, learning), and what I refuse to treat as a scoreboard (branded goods, luxury signaling, status upgrades).

2. Social media and advertising hygiene: I rarely swipe on social media; when I do, it’s mainly to keep up with close friends I already speak to in real life. I follow a few finance-related, tech-focused profiles and other interests, but I don’t use social media as a lifestyle feed. Once a year, I do a spring cleaning of who I follow on Facebook and Instagram. In parallel, I treat advertising as background pressure: I unsubscribe from retail and promo e-mails, turn off marketing notifications, use ad-blocking where practical, default to a privacy browser on mobile, and avoid browsing that’s really just “shopping for a feeling”.

3. I don’t finance wants: I treat credit as a payment tool and safety tool, not a way to afford upgrades. I don’t put purchases on installment plans or BNPL. If I wouldn’t pay in full today, I assume I can’t afford the upgrade, and I pass.

4. Make money talk normal and useful at home: I talk about money openly with my family and friends. There’s no taboo; I keep it relevant to my life and focus on principles rather than numbers - budgeting, saving, investing, long-term planning, and the trade-offs behind choices. The goal is to learn from each other and build financial discipline as a shared language, so good decisions don’t depend on secrecy or last-minute stress.

Final Thoughts

Consumer culture sets the default, and the default is spend. In economies where consumer spending is structurally central, and advertising, credit, and social comparison are always on, the upgrade-to-obligation pipeline isn’t a personal moral failure, it’s the path of least resistance.

The way out is to build a small counterculture around yourself. Decide what “enough” looks like, control what gets your attention, refuse to finance upgrades into obligations, and lock in surplus before it gets absorbed by a higher baseline. That’s how slack stops being an accident and becomes something you protect on purpose. Once slack is protected, everything else gets easier: shocks stay small, choices stay flexible, and “living well” stops depending on keeping up.

For many households that could be saving more, the pattern is rarely “too many lattes”. It’s recurring commitments that feel rational in the moment, small “yeses” that compound into fixed costs, and long-horizon goals that never feel urgent. The question isn’t “why don’t people care about the future?” It’s why spending feels vivid now while saving feels abstract.

Before the “how”, a reminder from the Lifestyle Inflation Trap: slack (income left after fixed costs) is the margin that lets you absorb shocks, invest, and make choices without panic. This piece is about why we leak that margin, psychologically and structurally, until it disappears.

Instant Gratification and Brain Chemistry: Spending Is a Fast Reward

Spending can feel rewarding in the most literal sense: it engages the brain’s reward and motivation circuitry. A key neuroscience point is that dopamine is more strongly tied to “wanting” (anticipation and pursuit) than to “liking” (the actual enjoyment after the purchase), which helps explain why the chase can matter more than the object. Once the new purchase becomes the new normal, the urge often returns.

Modern shopping environments intensify anticipation: novelty, deals, limited drops, and algorithmic feeds keep “wanting” activated. Some psychology and behavioral-health writing notes the same practical pattern: shopping can deliver a short-lived mood boost by triggering “feel good” chemicals, but it’s easy to overuse it as a coping strategy.

At the extreme end, researchers study compulsive buying as a clinical problem with ~5.8% of the US population affected over a lifetime. Most readers won’t relate to the clinical end of the spectrum, but the underlying loop (anticipation, brief relief, then regret and a new urge) shows up in milder form for a lot of people.

Dopamine rarely ruins finances through one splurge; it does damage when “small upgrades” become repeated behavior or recurring bills that pre-commit future income. Delay the purchase until the “wanting” cools. For non-essentials, wait 24-48 hours and re-check: “Would I still want this if I had to pay in full today?” If yes, it’s a choice; if no, it was craving.

Present Bias: “Future You” Is Abstract, “Now” Is Real

Even without dopamine-driven temptation, humans are predictably biased toward the present. Behavioral economists call it present bias (often modeled as “hyperbolic discounting”): we value immediate rewards disproportionately more than future rewards, even when the future payoff is larger. This is a cognitive bias rooted in our evolutionary history - survival often depended on immediate gains, not far-future planning.

That’s one reason saving and investing feel emotionally “quiet” day-to-day: the benefits are delayed, compounding, and mostly invisible. This also shows up in self-control research: delaying gratification is hard for many people, and early work popularized this in the “marshmallow” paradigm - though later research suggests outcomes depend heavily on context and constraints, not just “willpower”.

Regulators and behavioral researchers repeatedly note that short-term pressures and short-term “wants” can dominate long-horizon decisions, especially when planning is complex or stressful. Bridging the gap to your future self can help - e.g. making long-term outcomes more concrete through projections and simulations.

Present bias is how optional upgrades become fixed commitments, because the cost is spread into “manageable” monthly payments while the long-term trade-off stays intangible. Convert long-horizon saving into a monthly bill. Automate a transfer on payday, and increase it whenever income rises. If you rely on “what’s left at month-end”, present bias wins. Make “future you” concrete - run a simple projection once a quarter so the trade-off is visible, not abstract.

Social Comparison and Status Signaling: “Keeping Up” Is a Financial Force

A lot of spending is social. “Keeping up with the Joneses” is a description of how humans calibrate “normal”. If your peer group upgrades housing, cars, schools, vacations, and wardrobes, those choices can stop feeling optional and start feeling like baseline membership fees. Economists and sociologists have described conspicuous consumption for over a century; Thorstein Veblen’s core idea is that some purchases function as status signals, not just utility. The labels have changed, but the incentives remain.

This ties to evolutionary psychology: throughout history, displaying wealth could signal genetic fitness or social power, aiding in mating and social competition. Modern work adds a biological channel: an experiment found that testosterone administration increased men’s preference for status goods/brands, consistent with the idea that dominance-linked motives can increase “rank display” behavior. The “Big Three” are identity-linked and hard to downgrade. Once you buy into a neighborhood, a school path, or a car tier, stepping down can feel like a loss of status even if the upgrade didn’t meaningfully improve long-term well-being.

The bottom line is, the need for external validation and peer approval is a profound motivator for spending. It requires self-awareness and confidence to resist that and live frugally when everyone else seems to be upgrading.

Social comparison is how upgrades become defaults - because “normal” rises to your reference group, and monthly commitments lock that normal in. Don’t outsource “normal” to peers: Decide what you won’t compete on (zip code, car tier, brands). Make that decision once, not repeatedly at point of purchase.

Subscribe now

Emotional Spending: “Retail Therapy” Is Real

Not all spending is about status or novelty; some of it is emotion regulation. Stress, sadness, boredom, or anxiety can trigger shopping because it can create a sense of control and a short-term mood lift - an effect supported by consumer-psychology research on “control restoration” and buying decisions.

The issue is when shopping becomes a habitual coping mechanism. Many clinicians and researchers describe a similar cycle: shopping provides quick relief, the relief fades, and the original stressor is still there - sometimes with guilt or debt added on top. This is why simplistic advice often fails. If spending is serving an emotional function, “just budget harder” can miss the point. The spending isn’t only economic; it’s self-regulation.

Emotional spending doesn’t just cost money; it can convert mood volatility into financial volatility, increasing reliance on debt when real shocks hit. Pre-commit a “stress protocol”: When you feel the urge to buy for relief, do a short substitute first (talk to someone, walk, workout, shower). If you still want the item after the cool-down, it’s a considered purchase, not self-medication.

Low Financial Literacy and Weak Planning Habits

A straightforward contributor is low financial literacy: many people were never taught compounding, inflation, debt math, or basic budgeting. Large-scale surveys repeatedly find that financial literacy is associated with better planning and wealth outcomes.

Without a planning habit, money follows the path of least resistance: spend what’s left. If you don’t set clear saving goals or limits on spending, it’s easy to let extra money slip away. Moreover, without basic financial literacy, individuals may not understand the peril of high-interest debt or the benefits of starting to invest early.

In Italy, multiple sources place financial literacy meaningfully below many peers; one Bank of Italy paper cites international survey evidence around the high-30% range for “financially literate” adults. In the UK, the FCA’s Financial Lives research program repeatedly highlights gaps in resilience, confidence, and planning capacity, especially under cost-of-living pressure, creating a setting where present bias and easy credit do the rest.

Low literacy doesn’t just reduce returns; it increases the odds that surplus turns into consumption because there’s no default system capturing it. Use a “one-page” system, not motivation: One automatic savings/investing transfer, one spending account, one weekly 10-minute check. The point is defaults: you shouldn’t need to remember to be responsible.

Share Level-Headed Investing

Putting It Together: The Predictable Leak

All humans are somewhat wired to value the present over the future, seek pleasure, and care what others think, which is why long-term financial discipline is hard. These forces stack: present bias makes “later” feel optional. Dopamine makes “now” feel rewarding. Social comparison defines what “normal” costs. Emotional spending adds volatility. Low financial literacy and weak habits remove the system that could compensate for all of the above. In that environment, the default outcome of a pay raise is often not more stability and freedom, but a more expensive baseline.

This is the lifestyle inflation trap in behavioral terms: adaptation pushes upgrades; credit makes upgrades easy to convert into monthly payments; and monthly payments reduce slack until financial fragility appears.

Final Thoughts

If you accept these forces, the goal stops being “be disciplined” and becomes: design conditions where discipline is less necessary. Make saving the default outcome, not the heroic choice.

Automation beats willpower. “Pay yourself first” works because it captures slack before lifestyle adjusts (and before present bias gets a vote). Shrink the decision surface area: friction, smart defaults, and reversible luxuries reduce how often you have to fight your own brain - especially in environments where revolving credit and BNPL can turn a momentary want into a long-tail obligation.

Finally, match the tool to the constraint. As I laid out in The Demographics of Slack and Saving the right lever depends on where you sit in the slack distribution and what’s doing the squeezing. If the squeeze is obligations (especially the “Big Three”), avoid irreversible upgrades and unwind commitments. If you have surplus, automation, friction, and visibility do most of the work.

If there’s one meta-lesson from behavioral research, it’s that self-control is a scarce resource. People do better when the environment does more of the work.

I focus on Italy, the UK, and the US because they reflect different mixes of wages, housing costs, safety nets, and consumer credit - and I’ve lived in two and invested primarily in the third. Terms are defined in the Lifestyle Inflation Trap; here I use them without repeating the definitions.

Young Adults: Gen Z and Millennials, the “Soft-Saving”, Rent-Squeezed Generations

Gen Z (1997-2012) and Millennials (1981-1996) are shaped by high housing costs, weaker early-career income, and persistent uncertainty. In surveys, many Gen Z respondents describe a “soft saving” posture: prioritizing quality of life and experiences today over long-term accumulation.

Housing is the biggest expense, and it shows up as a stark age gap in housing cost burden (the share of income absorbed by housing). In the US, households headed by 25-34-year-olds spend 30% of pre-tax income on housing, versus 17% for 65+ households, who more often benefit from paid-off mortgages. In the UK, renters aged 25-34 spend 31% of income on rent, while 65+ homeowners spend a tenth. In Italy, younger adults who form independent households spend 31.6% of income on housing, while the 65+ cohort spends 15% on maintenance and utilities.

Once housing eats the margin, everything else becomes downstream. In the UK, ages 25-34 post the lowest saving ratio of any demographic, and a large minority of under-35s report having less than £1,000 in cash savings. In Italy, the demographic split is unusual: low entry-level wages and precarious contracts keep between 64.3% and 71.5% of adults aged 18-34 living with their parents (vs. EU’s average 49.4%). That arrangement functions as a “forced” safety net and it reduces visible hardship but delays independent household formation. In the US, Millennials carry the highest levels of student loans and credit card debt, and despite being in prime earning years, about 42% of Gen Z and Millennials live paycheck-to-paycheck (P2P) as of October 2025.

For younger cohorts, housing is the binding constraint - it compresses slack before lifestyle choices even enter the picture.

The Middle-Age Paradox: More Money, Less Liquidity

Middle-aged adults (Gen X, 1965-1980 and younger Boomers) sit at the intersection of peak earning power and peak financial obligations. In the US, household income peaks around ages 45-54, and that’s also when many households take on their largest, stickiest commitments.

This is the decade when the “Big Three” often peak at the same time. The culturally reinforced “successful adult” script (bigger home, better car, more comfort and convenience) can convert rising incomes into higher “defaults”, so liquidity doesn’t rise in step with income.

The data points line up with the pattern. In the US, Goldman Sachs’ 2025 Retirement Survey reports that 40% of people earning above $300,000/year say they live paycheck-to-paycheck, consistent with fixed commitments eroding slack even at high incomes.

In the UK, income tends to peak in the 40s, and expenditures are also highest in prime working ages, when mortgages/rents and family costs are most intense. In Italy, household expenditure is higher for working-age adults and higher-status professions than for older households; for example, a single person aged 35-64 spends €2,096/month versus €1,794/month for 65+, and managers/professionals average €3,386/month.

In the US, healthcare exposure and childcare can turn midlife into a peak-obligation period even at high incomes; in the UK and Italy, healthcare is not a private-cost shock, but housing and family costs still dominate the monthly baseline. The risk is locking peak income into peak fixed commitments.

In midlife, obligations peak alongside income, so even strong income can coexist with low liquidity if “defaults” expand with them.

Share Level-Headed Investing

Silver Wealth: Boomers and the Spending Power of Assets

Baby Boomers (1946-1964) sit in a different financial regime: their spending power is often driven more by accumulated assets than by current wages. In the US, Federal Reserve-based summaries put Boomers at 51% of total household wealth, far ahead of Gen X and Millennials, which helps explain why “silver spending” can stay resilient (travel, experiences, premium/luxury) even when younger cohorts retrench.

The key driver is balance sheet strength. In the US, older households are more likely to own homes, and decades of home-price appreciation converted ownership into net worth (often via paid-off mortgages and rising equity). They also benefited from an earlier retirement landscape where defined-benefit pensions were more common, before the shift toward 401(k)/defined-contribution plans pushed more retirement risk onto individuals.

The picture has an Italian analog. Home ownership is especially concentrated among older cohorts (76% of seniors own their homes), and historically higher pension replacement rates have helped stabilize retirement income. Research describes over-65s as being in a better economic condition than other age groups, reflecting the cumulative results of lifetime work, saving, investing, and real-asset ownership - often paired with lower day-to-day spending propensity, but meaningful capacity for discretionary spending when desired.

The constraint shifts from paycheck to balance sheet; assets (housing equity and pensions) can sustain discretionary spending even when younger cohorts are income-squeezed.

“HENRYs” – High Earners, Not Rich Yet

High Earners, Not Rich Yet (“HENRYs”) are households with high incomes but limited wealth, because spending and fixed commitments tend to scale up alongside income. The label is often used for $250k-$500k households in the US and £100k+ earners in the UK, and it’s common in high-cost cities where the “baseline” for housing, schooling, and lifestyle is already expensive. Goldman Sachs’ analysis notes that 32% of US HENRYs report living P2P.

The issue is that high income doesn’t automatically translate into slack when big-ticket decisions become structural. Upscale rent or a larger mortgage, premium childcare/schooling, car leases, and a higher “default” level of convenience can pre-commit a large share of monthly cash flow. On paper, income looks strong; in practice, liquidity can stay thin because obligations rise faster than savings.

Layered on top is status consumption. For the top 10% of earners, a larger share of spending shifts toward “Veblen goods”, luxury items where price itself signals status. Research suggests this cohort accounts for 49% of total US consumer spending, driven heavily by housing, luxury transport, and designer brands -categories that can be both identity-linked and hard to downgrade once adopted.

High income can still be fragile when “success” gets converted into fixed monthly commitments faster than wealth is built.

Lower Income and The Financial Struggle

At the bottom of the income distribution, the binding constraint is arithmetic, not “discipline”: essentials consume most of the paycheck, so slack is close to zero and many households live month to month.

This shows up as paycheck-to-paycheck being driven by necessity spending. One widely used Bank of America Institute definition classifies a household as P2P when “necessity spending” exceeds 95% of income, capturing what “making ends meet” looks like for 24% of low earners, even with disciplined spending.

The consequence is “dis-saving”. In both the US and UK, the poorest households often spend more than they bring in, covering the gap by drawing down savings, borrowing, or falling behind on bills. In the UK, Citizens Advice estimates over 2.1 million households (7.67% of the population) in England and Wales are in a “negative budget” (income too low to cover bare essentials), which predictably pushes people toward arrears and debt. In the US, BLS work finds personal saving is negative for the bottom half of the income distribution, and for the bottom 10% expenditures can be roughly double income.

Even under tight budgets, status pressures can distort choices and visible purchases can sometimes serve social-signaling or coping functions, though the primary driver of fragility remains essentials absorbing most income.

When slack is near zero, “cut spending” has limited leverage - financial fragility is largely structural, and any shock quickly becomes debt.

Subscribe now

Middle-Class Squeeze vs. Lifestyle Creep

A meaningful share of P2P pressure in the middle class isn’t just lifestyle creep; it’s cost-of-living squeeze. The baseline cost of “normal” middle-class life has risen faster than incomes in many places, so era matters, not just age.

The story is fixed costs: housing first, then big-ticket essentials like healthcare (US) and childcare (UK). When these rise faster than income, slack shrinks before any discretionary choices enter the picture, and the same lifestyle can become fragile.

In the US, real median household income1 rose 16.6% from $71,790 (2000) to $83,730 (2024), while house prices rose 211% and healthcare spending has tripled - an imbalance that compresses slack even without upgrades.

In the UK, affordability deteriorated sharply: in 2002 the average home cost about 5x average gross household income; today it’s about 8.2x - levels of unaffordability not seen since the Victorian era.

In Italy, the issue is income stagnation: OECD highlights real wage decline over the past two decades, and Eurostat data confirms real household income per capita2 declined by 4% over 2004-2024, meaning essentials bite harder with less income momentum.

When the middle-class baseline absorbs the margin, lifestyle creep becomes more dangerous: with less slack, optional upgrades or small recurring commitments can tip budgets from tight to fragile.

The Squeeze vs. The Creep: What The Demographics Say

In summary, lifestyle creep is widespread across age and income groups, but it manifests differently. Younger adults often tilt spending toward experiences and immediate gratification, delaying long-horizon saving. Midlife and high earners are more likely to convert rising income into larger fixed commitments - sometimes to the point of financial fragility despite strong salaries. At the bottom of the income distribution, the binding constraint is usually arithmetic: essentials absorb most income, leaving little room to save. In other words, without deliberate restraint, a higher salary can simply finance a higher “default” lifestyle rather than translate into higher savings and long-term financial security.

A few reality checks matter. Some spending that looks like lifestyle creep is a rational trade-off: young families may pay more for safer neighborhoods, shorter commutes, or better schools. Cross-country comparisons also require humility - where public pensions, healthcare, and safety nets are stronger, households may rationally rely less on private saving. Not everyone is optimizing for net worth; some prioritize experiences and present quality of life, which can be reasonable if it doesn’t depend on debt or perpetual income growth. And for many households, the constraint is arithmetic, not behavior: fixed costs have eaten the margin. These nuances don’t contradict the lifestyle inflation story - they clarify the line between unavoidable squeeze and avoidable upgrades.

Share

Final Thoughts

The most useful next step isn’t “be more disciplined”. It’s to diagnose what’s actually squeezing you, because different problems need different tools. Put your situation in one of three buckets.

If it’s arithmetic, essentials absorb the paycheck. Housing, utilities, food, transport, and basic bills leave little margin. In that case the priority is stability: avoid high-interest debt, build even a small emergency buffer, and don’t add new fixed costs. The highest-leverage moves are structural (income, benefits, housing choices, shared living, cheaper transport), not fine-grained budgeting - and one of the most durable levers on income is investing in yourself: building skills, credentials, and experience that let you progress in your current career or pivot to a better-paying role.

If it’s obligations, your problem isn’t low income; it’s contracts. A mortgage/rent at the edge, car payments, childcare/schooling, installments, and subscriptions have pre-committed most of your future cash flow. Here the goal is to stop hardening your baseline further and to unwind what you can. Treat every new monthly payment as a veto on future financial flexibility. Refinance where possible, eliminate “stealth” recurring costs, and pressure-test the “Big Three” before you upgrade anything.

If it’s upgrades, you do have slack, but it leaks into lifestyle defaults. This is where lifestyle creep is most dangerous because it feels harmless: small “yeses” that become “normal”. Your solution is default design: capture raises automatically (“pay yourself first”), add friction to impulse spending, and keep luxuries reversible. Make saving the path of least resistance so you’re not relying on willpower in an environment built to sell you convenience.

Once you name the squeeze, the right playbook becomes obvious. Arithmetic calls for stabilization. Obligations call for de-commitment. Upgrades call for guardrails. The mistake is applying the same generic advice to all three.

Since 2009, my income has nearly quadrupled, and the main reason my spending didn’t mechanically follow is that I tracked it closely and put guardrails in place early. My fixed costs still rose with family life and later fell again, while I lived below my means and tried to keep upgrades reversible and avoid letting new “defaults” turn into permanent monthly bills. That personal experiment is the lens for what follows.

I focus on Italy (where I live), the UK (where I spent 13 years), and the US (where most of my portfolio exposure sits) because they’re familiar to me and because they represent three different mixes of wages, housing costs, welfare support, and consumer credit that shape how easily spending turns into fixed commitments.

This isn’t a moral judgment - plenty of households are constrained by housing, childcare, and wage stagnation, with fixed costs leaving little room to save. This article focuses on what happens when surplus exists and quietly turns into recurring commitments. Next, I’ll define the key terms and lay out a simple dashboard to spot where slack is leaking.

The Slack Framework (Definitions + Indicators)

Slack is the income left after fixed costs - your “margin”. It’s what lets you absorb surprises, save, invest, and make decisions without panic.

Why slack matters: Saving is the buffer that prevents ordinary problems from becoming financial emergencies. A cash reserve turns a car repair, medical bill, or job gap into a manageable setback, instead of a debt spiral. Investing comes after that foundation. Once an emergency buffer exists, investing is what you do with surplus slack so it can outpace price inflation1 and compound2 over time.

Lifestyle inflation (or lifestyle creep) is what happens when rising income turns into higher “defaults”, especially fixed monthly commitments, so your lifestyle expands without you explicitly choosing it.

The “Big Three” are the high-impact categories where upgrades most often become sticky: housing, transportation, and education/childcare.

Fixed-cost creep is the slower leak: subscriptions, app add-ons, device financing, installments, premium tiers - small recurring payments that compound into a meaningful monthly drain.

Because slack isn’t reported directly in national statistics, I use a small dashboard of indicators as proxies for household financial fragility. When saving rates3 are low, it suggests less income is being retained as an emergency buffer. When more people report living paycheck-to-paycheck4, it signals tighter month-to-month cash flow. When emergency buffer gaps are large, it implies many households can’t absorb a routine shock without borrowing. Finally, the Debt Service Ratio (DSR)5 shows how much income is pre-committed to lenders (often the downstream result of upgrades that became contractual obligations), before groceries, rent, childcare, or savings.

Definitions differ across sources (macro accounts, central banks, surveys); levels are indicative. The point is direction and pressure (slack), not precise cross-country ranking.

Together, the indicators explain why lifestyle creep is easy to fall into: as prices rise and credit stays available, spending can become structural. Price inflation also resets what “normal” costs, and people adapt quickly - so the baseline drifts upward even without deliberate upgrades.

Where the Trap Springs: The “Big Three” + Fixed-Cost Creep

Financial behavior across Italy, the US, and the UK isn’t just about “discipline” or personal choices; it’s shaped by slack, what’s left after fixed costs. Over the past two decades, slack has narrowed for many younger households and for the “squeezed” middle class as housing and other essentials take a larger bite. But when surplus does exist, the trap usually isn’t daily coffee. It’s the conversion of liquid cash into sticky, high-status commitments: the “Big Three”.

The “Big Three” are the headline drivers because they’re identity-linked and hard to downgrade. Housing upgrades (a larger mortgage or pricier neighborhood), cars/transport upgrades (buying or leasing high-end vehicles that signal status but depreciate rapidly), and education/childcare upgrades (after-school, tutoring, private schooling, premium extra-curricular) don’t just raise spending, they harden it into fixed monthly outflows. This is where cost-of-living “squeeze” (rising rents, bills, essentials) ends and lifestyle creep begins: voluntary upgrades that become the new baseline. For high earners in particular, these categories inflate disproportionately because they’re tied to identity and status - once you upgrade your zip code, car, or schooling choices, downgrading can feel like a loss of social standing, turning what was once a luxury into a permanent cost.

There’s also a quieter form of lifestyle creep that rarely feels like “spending”: fixed-cost creep. Modern consumption increasingly arrives as a bundle of recurring payments - subscriptions (streaming, cloud storage, gyms), app add-ons, device financing, and “Buy Now, Pay Later” (BNPL)6/installment plans that turn discretionary wants into recurring obligations. Each one is small and easy to justify in isolation (“it’s only $15/month”), but together they compress slack the same way a bigger mortgage does: they pre-commit tomorrow’s income before you’ve even made choices about saving, investing, or experiences. And because these costs are fragmented and often auto-billed, they’re easy to ignore - until a shock hits, or your income dips and you discover how little room you actually have.

A simple way to see it: $15/month doesn’t feel like much, until it’s 12 different subscriptions ($180/month, $2,160/year). Add a couple of 0% installments (phone, laptop, furniture) and a lease upgrade, and you’ve built a stealth mortgage: a stack of contractual monthly outflows that behaves like fixed costs, even though it started as optional.

This is where lifestyle inflation really happens - not in one big splurge, but in commitments that quietly turn your future slack into someone else’s recurring revenue.

The Hedonic Treadmill: Why More Becomes the New Normal

One of the strongest drivers of lifestyle creep is the hedonic treadmill (hedonic adaptation): we quickly get used to improvements (a pay raise, a new car, a bigger house) and drift back toward a baseline level of happiness. What starts as a treat becomes “normal”, expectations rise, and you start chasing the next upgrade just to get the same buzz.

Lifestyle creep is the treadmill playing out in your budget: higher earnings fuel higher “defaults”, and the real trap isn’t the upgrade itself, it’s when the upgrade becomes a recurring bill. The pleasure fades, but the bill stays, shrinking slack month after month. This is how higher income turns into higher defaults: adaptation pushes you to upgrade, and credit makes it easy to lock the upgrade into monthly payments, reducing slack.

The practical takeaway is simple: if you let temporary treats become permanent obligations, your “normal” gets more expensive faster than your freedom grows. Even if the boost fades within weeks or months, the higher fixed costs remain - consistent with adaptation findings since the classic 1978 study and later work pointing in the same direction.

Share Level-Headed Investing

The Dopamine Loop: How Wanting Becomes a Monthly Bill

The biological engine of the hedonic treadmill is dopamine. Neuroscience separates “wanting” (craving/anticipation) from “liking” (actual enjoyment). When you want a new purchase, dopamine fuels anticipation; after you buy it, that spike fades, the item becomes the new “normal”, and the brain starts looking for the next hit. Marketing and social feeds exploit this loop by constantly resetting what a “good life” looks like, so lifestyle creep can feel automatic: your brain rewards pursuit more than possession.

A second force is present bias (hyperbolic discounting): we discount the future steeply, so immediate rewards feel disproportionately valuable. That’s why a smaller reward now can feel “worth more” than a larger reward later. In practice, this nudges people toward upgrades today and saving (and investing, if you do it) later, until the upgrade hardens into a fixed monthly commitment and slack quietly shrinks.

I can’t turn these instincts off, so I try to design around them. I steer discretionary spending toward experiences (travel, meals out) rather than possessions, because experiences become memories and part of your life story and identity while objects fade into the background - consistent with consumer-psychology evidence that doing tends to outlast having. As my income rose, I worked to keep luxuries optional rather than default, and I automatically route a meaningful share of raises/bonuses into investments before my baseline adjusts, so the default outcome of a raise isn’t a new recurring bill.

Final Thoughts

Lifestyle inflation rarely arrives as one big splurge. It arrives as “defaults”: upgrades that become fixed monthly commitments and quietly convert future slack into recurring bills. If you take nothing else from this, take the mechanism seriously - once a cost becomes contractual, it stops being a choice and starts acting like a tax on every future paycheck. I’m aware this lands differently depending on your starting point: if you’re squeezed, the goal is survival and stability (protecting slack and avoiding debt); if you have surplus, the goal is discipline, so pay raises don’t quietly harden into permanent commitments.

What’s helped me is treating slack like an asset to protect, not leftover cash to spend. My personal guardrails:

1. Capture raises automatically (“pay myself first”) - when income jumps, I immediately increase the amount that’s auto-routed to my investing account. I treat it like a bill I owe my future self, so the raise gets captured before my spending baseline resets.

2. Audit the “Big Three”: as my income grew, I deliberately kept housing from expanding with it - so the share of my income going to rent (early years) and then a mortgage (recent years) fell from 47% in 2009 to 21% on average over the past five years. The goal isn’t austerity; it’s avoiding permanent commitments that shrink slack.

3. Add friction to impulse buys: I use a 24-hour rule for non-essentials, avoid one-click checkout, and run a quick “cool-off” period before purchasing.

4. Prefer experiences and relationships over status (for readers with surplus slack, not for households in financial fragility): I’m happy spending on travel and meals out (especially when shared with people I care about), rather than purchases meant to signal status. As my income grew, I increased travel/experience spending from 6.2% of my income (2009-2013) to 14% (2021-2025).

If you want one practical starting point, measure the two places slack leaks fastest, and make the measurement a habit. First, compute your “Big Three” as a share of income and track it monthly for a year. Second, list every recurring payment (subscriptions, app add-ons, BNPL/installments, device financing), sum them, and annualize the total. This is the real value of tracking expenses and budgeting: not micromanaging every coffee, but making the invisible visible so “stealth fixed costs” don’t quietly become your new baseline. Seeing the one-line yearly number is often the fastest way to spot what you stopped noticing.

Since the March 2025 IPO, the CoreWeave () stock has lived that tension in real time - sharp post-IPO upside, then a violent repricing as investors started focusing on funding costs and execution risk rather than growth slogans.

Q3 2025 in One Sentence: Revenue Doubled, Operating Income Fell

On the headline numbers, Q3 2025 was spectacular: revenue was $1.4 billion, up 134% year-over-year (YoY), and backlog was $55.6 billion.

But the income statement tells the less-market-friendly version: operating expenses also exploded to $1.3 billion (+181% YoY), leaving GAAP operating income at $51.9 million, down 55.6% from $117 million a year ago: more growth but less operating income.

This is the Buffett line in numbers: “not all growth is good growth”, especially when the expense structure (and financing structure) grows faster than the underlying profitability.

In my original piece, I flagged the trajectory: rising debt would drag interest expense from large to structural, and eventually force the business to run for its lenders before it can run for shareholders. Q3 2025 makes that concern concrete: Interest expense was $310.6 million in Q3, nearly 3x the $104.4 million in Q3 2024. As a share of revenue, that’s about 23%. For the first nine months of 2025, interest expense was $841.4 million, up 297% versus the prior year period. Cash paid for interest (net of capitalized amounts) was $557.1 million year-to-date (YTD), versus $97.4 million the year before; showing this isn’t just an accounting artifact.

This is exactly what “growth funded by debt” looks like once rates and borrowing levels catch up: the financing line starts consuming the income statement.

Adjusted EBITDA: The Clean-Looking Metric That Removes the Messy Reality

CoreWeave’s Q3 2025 earnings deck leads with $838 million of adjusted EBITDA (61% margin). The problem is that for a business where assets age quickly and the fleet must be refreshed continuously to stay competitive, depreciation is economically closer to a recurring reinvestment requirement than a disposable, non-cash nuisance; precisely the point I made in the original article.

In Q3, depreciation and amortization on property & equipment was $630.5 million for the quarter. Add that to the $310.6 million of interest expense, and you get $941 million, about 68.9% of quarterly revenue; before you even factor in taxes.

When those two lines are treated as “below the line” and effectively ignored in the headline narrative, the business can look far more profitable than it is.

Cash Flow and CapEx: the Business is Still Not Self-Funding

CoreWeave generated $1.5 billion of operating cash flow (OCF) in the first nine months of 2025, down from $2.6 billion the year before. It spent $6.25 billion on purchases of property & equipment (CapEx) over the same period. The business model remains the same: operating cash helps, but the expansion plan overwhelms it, producing deeply negative free cash flow that must be bridged with debt. YTD proceeds from issuance of debt were $7.6 billion, with $3 billion of repayments. The company also discloses in their Q3 10-Q filing that debt-service cash flows totaled $3.6 billion YTD, including $3 billion of principal payments and $651 million of interest payments (including capitalized interest).

Balance Sheet: the Refinancing Question Didn’t Go Away, It Moved Closer

At September 30, 2025, total current assets were $4.7 billion, while total current liabilities were $9.7 billion, about a $5 billion gap, with $3.7 billion of current debt inside those current liabilities. Total debt (current + non-current) was $14 billion at quarter-end. This compares uncomfortably with the framing from my original article: a business whose “growth engine” requires external capital staying continuously available, at tolerable rates, while the asset base depreciates rapidly.

The “more debt” drumbeat hasn’t stopped after quarter-end: in December 2025, CoreWeave priced an upsized $2.25 billion convertible senior notes offering.

Execution Risk Meets Leverage: Capacity Delays Are Not Harmless in This Setup

During the Q3 earnings call, CoreWeave trimmed its outlook due to delays at a third-party data center partner. In a lightly levered model, delays are annoying. In a heavily levered, high-CapEx model, delays can become financially meaningful because you keep paying interest while the revenue you expected to support that interest arrives later.

CoreWeave announced an all-stock deal to acquire its data center partner Core Scientific in July 2025, explicitly tying it to scale, control, and cost of capital. But Core Scientific () shareholders ultimately voted it down and the deal was terminated in late October 2025, leaving CoreWeave still exposed to partner execution while maintaining the same broad leverage posture.

Share

What Changed Since My Original Article? The Numbers Caught Up

If you strip this follow-up down to the thesis test:

• Debt has continued to expand - $14 billion of debt on the balance sheet at September 30, 2025, plus additional post-quarter financing activity.

• Interest costs are rising fast and have become a major share of revenue - 23% in Q3.

• The company’s preferred profitability framing still leans on adjusted metrics that, by design, step over depreciation and financing - the two expenses most tightly coupled to long-term economics in this kind of business.

Final Thoughts: Obvious Growth Still Isn’t Obvious Returns

AI infrastructure demand can be real and enormous, and still be a tough place for shareholders if the industry’s returns get competed away while financing costs compound. CoreWeave’s Q3 2025 filing makes that trade-off sharper, not softer: faster scaling, heavier capital intensity, higher interest burden, and more reliance on continuous access to capital markets.

I am not a CoreWeave’s shareholder, and I do not intend to become one. More broadly, I have no interest in owning “GPU renters” as a category. As I argued in my original article, I struggle to see an intrinsic moat in a business model where the core assets are purchasable by anyone with capital, the product is largely a commodity (compute), and differentiation risks collapsing into price, financing access, and execution. In that setup, even strong demand can translate into mediocre investor economics.

The next few quarters, I’d watch three things more than revenue figure and adjusted EBITDA: interest expense and cash interest paid, net debt growth versus incremental gross profit, and whether CapEx intensity falls fast enough to credibly point toward durable free cash flow, without requiring one more major debt raise.

Management is not subtle about the scale. In Q3 2025, Amazon reported capital expenditures (CapEx) of $34.2 billion, and $89.9 billion year-to-date (YTD), and explicitly expects these investments to rise further into 2026.

The payoff isn’t one AI product. It’s margin expansion across multiple lines: fulfillment and delivery cost-to-serve, advertising monetization, subscription retention, and AWS gross profit dollars. Over a long horizon, Amazon’s strongest pattern has been to build infrastructure once, then monetize it many times over.

Subscribe now

Commerce Margin Through AI and Robotics

The cleanest long-term retail thesis for Amazon is not e-commerce growth. It’s cost per package down, speed up, and conversion up, enabled by automation.

In June 2025, Amazon’s warehouse robotics fleet surpassed one million units across 300 facilities, and the company has continued to push automation deeper into more complex tasks. In May 2025, Amazon unveiled Vulcan, a robot that combines sight and touch to handle items in human-like environments. Then in October 2025, Amazon introduced Blue Jay, which coordinates multiple robotic arms to pick, stow, and consolidate items in a single workstation, collapsing what previously required three separate robotic stations into one. On top of the physical robotics layer, Amazon is building an AI control plane for the network. DeepFleet coordinates and optimizes robot movements to reduce travel bottlenecks, improving robot efficiency and travel time by 10%. In addition, Amazon is layering in agentic AI for operators as well: Project Eluna is designed to cut through dozens of dashboards by using real-time and historical building data to anticipate bottlenecks and recommend actions to keep operations flowing smoothly.

Taken together with AI-powered inventory forecasting, reported to have improved regional forecast accuracy by 20%, these advances point to the same outcome: fewer wasted movements, faster throughput, and a structurally lower unit cost to fulfill and deliver packages. Importantly, Amazon’s fulfillment centers also make clear that the end-state is smarter human-robot collaboration. Robots handle repeatable or ergonomically difficult work (for example, Vulcan can pick from high and low shelves), while people stay focused on the edge cases, maintenance, and judgment calls.

Over the past year, Amazon’s management has consistently emphasized automation, robotics, and AI integration in fulfillment centers as the mechanism to improve speed, unit economics, and profitability. In the Q3 2025 earnings call, CFO Brian Olsavsky made that explicit, while also setting expectations that the path won’t be linear:

Warehouse automation drives four reinforcing advantages:

1. Faster inventory handling, which improves delivery promises and conversion.

2. Fewer touches per item, which reduces variable cost.

3. Better inventory placement decisions, which reduces distance traveled and split shipments.

4. Higher reliability, which increases customer trust and purchase frequency.

Amazon repeatedly points to improvements in operations and enhancements to customer self-service features, which it attributes to investments in AI systems and automation - reducing service contacts per unit and improving cost efficiency. That matters because cost-to-serve is not just pick/pack/ship; it’s also customer care, returns workflows, issue resolution, and the countless small frictions that scale with volume.

The impact shows up most clearly in the North America segment, where Amazon connects automation and AI initiatives directly to profit margin expansion. Newer robotic systems, combined with a more regionalized network and algorithmic optimization, have improved throughput efficiency and reduced per-unit handling costs, supporting operating leverage even as the company continues to invest.

The 2026+ catalyst is that the automation curve is still early. Once you cross a critical mass of robots and standardized AI control layers, improvements become increasingly software-driven: better routing, better task allocation, better forecasting, better exception handling. That’s the highest-quality kind of margin expansion because it tends to persist rather than reset each year.

A practical watch item for 2026: North America operating margin. It won’t rise in a straight line, but the underlying drivers are structurally moving in the right direction.

Grocery & Perishables: The “Trojan Horse” For Frequency and Higher Lifetime Value

Grocery has low standalone margins, so it’s easy to misread it as a dead-end category. The better way to view grocery is as an engagement lever: it increases purchase frequency and makes Prime stickier. That stickiness then supports higher-margin attach (ads, subscriptions, services).

Its recent push into same-day perishables is a concrete, near-term execution milestone that extends into 2026. In December 2025, Amazon announced same-day delivery for fresh groceries expanded from 1,000 to 2,300+ cities and towns, with further expansion coming in 2026. CEO of Worldwide Amazon Stores, Doug Herrington, framed it as customers combining grocery with regular Amazon purchases “in ways that make their lives easier and save them valuable time”.

Two details matter:

• In same-day perishable markets, Amazon says “nine of the top ten best-selling items are now perishables”.

• Amazon says “customers who add fresh groceries to their Same-Day Delivery orders shop about twice as often as those who don’t”.

Frequency is the hidden asset. If grocery increases order cadence, Amazon gets more opportunities to monetize intent (ads), increase basket size, and reduce delivery cost per item through denser routes and better inventory positioning.

This is strategically important because rural America is exactly where incumbents like Walmart are strong. Amazon is trying to close the convenience gap, using its own last-mile build-out plus AI-driven local demand prediction to stock the right items closer to customers.

In its Q3 2025 10-Q filing, Amazon described “investments in our fulfillment network, partially offset by fulfillment network efficiencies”, which suggests early returns from AI-driven logistics and route optimization efforts.

2026 watch items:

• Expansion pace of perishables/same-day coverage.

• Evidence that grocery attach increases Prime retention and order frequency.

• Whether rural delivery improves unit economics (density, repeat purchase rates).

AWS’s 2026 Setup: Gigawatts + Trainium + AgentCore

The “AWS is slowing” narrative tends to appear when Microsoft Azure and Google Cloud Platform (GCP) post higher growth rates. The question I ask myself is simpler: will AWS keep capturing the AI workload wave profitably while maintaining strategic control over its cost base? AWS remains Amazon’s largest profit engine, and Q3 2025 was a good snapshot of why: revenue grew 20.2% year-over-year (YoY) to $33 billion while operating income increased 9% to $11.4 billion. CEO Andy Jassy framed the moment clearly:

From a 2026+ catalyst perspective, the core is capacity plus chips, with software increasingly acting as the multiplier. On the Q3 2025 earnings call, Jassy said AWS added 3.8 gigawatts of capacity in the last year, with “another gigawatt plus coming in the fourth quarter”, and expects to “double our overall capacity by the end of 2027”. The same discussion connects directly to custom silicon: AWS will keep buying NVIDIA and is “not constrained” in doing so; at the same time, it’s pushing hard on proprietary compute economics via Trainium (AI accelerators) and Graviton (CPUs).

What’s changed since earlier “AWS is falling behind in AI” narrative is that the custom silicon program now has tangible validation signals. Trainium2 adoption reportedly surged 150% quarter-over-quarter (QoQ) into a multi-billion-dollar business.

At re:Invent 2025 conference on early December 2025, AWS doubled down on the full-stack message - custom chips, foundation models, and distribution through AWS’s developer platform - highlighting Graviton5 and Trainium3 UltraServers alongside agentic infrastructure like Bedrock AgentCore, underscoring vertical integration as a core competitive advantage. Around the same time, SemiAnalysis published a deep dive positioning Trainium3 as a potentially serious challenger on cost/performance dynamics versus incumbent alternatives (NVIDIA’s GPUs and Google’s TPUs) - exactly the direction you’d expect if the AI market becomes more price-sensitive over time.

Jassy described Trainium2 as having “substantial traction, is a multi-billion-dollar revenue run rate business, has 1M+ chips in production, and 100K+ companies using it as the majority of Bedrock usage today”.

This fits with AWS’s strategy of anchoring custom silicon with large, demanding workloads: in addition to a million Trainium2 chips in production, Project Rainier connects more than 500,000 Trainium2 chips, built in collaboration with Anthropic to support Claude training and inference.

Jassy told investors Trainium3 volumes would be “much fuller” in early 2026 and linked adoption to cost-sensitive production inference workloads. In parallel, AWS is expanding its NVIDIA footprint with new Grace Blackwell instances, while continuing to broaden global infrastructure, adding major capacity, launching a region in New Zealand, and expanding its footprint with 10 additional Availability Zones.

This matters because it links directly to profitability. AWS operating margin improved, reaching 34.6% in Q3 2025 versus 30.3% two years prior, consistent with efficiency gains from proprietary silicon and scaling AI workloads.

At re:Invent 2025, AWS positioned Bedrock AgentCore as the production layer for agents. Jassy said it’s “worth double clicking”, emphasizing security and scalability, and highlighted new building blocks like Policy and Evaluations to control and measure agent behavior. He also argued “agents will become the primary way companies get value from AI”, pointing to AWS’s growing lineup (e.g. Kiro for coding) and to Nova Forge as a way to customize frontier models with proprietary data so companies can build differentiated agents, not generic wrappers.

The strategic advantage is that if AI becomes a larger share of global compute, the long-run winners won’t just rent GPUs. They’ll win by offering multiple price-performance tiers, integrating hardware and software, and controlling supply and economics through design and procurement. If AWS nails capacity availability, delivers compelling price-performance across both NVIDIA and custom silicon, and pairs it with developer-friendly agentic tooling, then the “AWS AI resurgence” story becomes about measurable throughput; accelerating growth and durable operating income dollars. That combination could also remove skepticism about AWS’s AI competitiveness - a major overhang that has likely contributed to the stock being roughly flat in 2025 - and set up a 2026 upside surprise as new capacity comes online and more workloads shift onto differentiated, lower-cost compute.

One more important signal is organizational: Amazon is reshaping its AI leadership structure, with SVP Peter DeSantis overseeing an organization spanning AI models, custom silicon (Graviton/Trainium), and quantum efforts.

2026 watch items:

• Capacity coming online fast enough to sustain >20% top line growth.

• AWS operating margin - does it keep widening as AI scales?

• Trainium3 adoption in real production inference.

• AgentCore moving from tooling to scaled production usage.

Advertising + Prime Video Profit Pool: High-Margin Attach on Top of Commerce

Amazon’s retail engine throws off something more valuable than gross merchandise volume (GMV): intent. When customers arrive on Amazon, they often arrive to buy. Advertising is the monetization layer on top of that intent, and Amazon’s scale is now material enough that modest improvements in ad yield can translate into large incremental operating profit.

In Q3 2025, Amazon reported ad revenue of $17.7 billion with growth accelerating to 24% YoY for the third consecutive quarter. Over 2025, management repeatedly framed the ad business as a high-margin layer synchronized with the retail ecosystem, built on engagement, attention, and data synergies across commerce and media.

Prime Video is part of the same ecosystem logic: it supports Prime value perception and creates more surfaces for monetization (subscriptions, advertising, and customer engagement). Jassy explicitly connected the dots, noting that video advertising is “already a very large amount of advertising revenue”, while “still relatively early stage”. By October 2025, Amazon broadened the advertising surface area well beyond its own storefront and video properties by expanding Amazon DSP’s reach through major third-party inventory and audio integrations, effectively turning DSP into a global distribution layer for Amazon’s ad demand. Jassy put it plainly:

On the owned-media side, he also highlighted that Prime Video live sports is drawing strong advertiser demand, with upfront commitments for 2025 and 2026 exceeding internal expectations.

Looking into 2026 and beyond, the combination of a scaled and increasingly efficient advertising engine with a fast-growing Prime Video ad platform positions Amazon with another multi‑billion‑dollar profit pool atop its core commerce base. In plain English: grocery increases frequency, Prime Video expands attention and perceived Prime value, and Amazon’s full-funnel ads monetize the entire journey, from discovery to checkout.

2026 watch items:

• Can ads sustain >20% growth as Prime Video/live sports scale.

• Does DSP’s broader reach convert into higher yield and bigger ad budgets.

Share Level-Headed Investing

Option Value That Could Become Real Value Beyond 2025

A long-horizon Amazon thesis should explicitly include options; projects that are not necessary for the base case, but can matter a lot if they hit.

• Amazon Nova: multimodal AI model family, introduced Nova Multimodal Embeddings and Web Grounding, expanding capabilities across text, image, and video search.

• Amazon Leo (“Project Kuiper”): In Q3 2025 earnings release, Amazon said it expanded the number of satellites to more than 150 and referenced 1+ Gigabit per second test speeds with an enterprise-grade terminal.

• Zoox: Amazon noted Zoox robotaxis are available to riders in Las Vegas and mentioned Washington, D.C. as the eighth testing location.

These projects highlight Amazon’s diversification beyond commerce and cloud and deepen its optionality in connectivity, mobility, and AI-enabled services; they won’t drive tomorrow earnings, but they are asymmetric bets where a successful scale-up into commercial platforms could deliver upside far beyond what today’s valuations reflect.

Key Risks to Respect

Risks matter because Amazon is scaling multiple capital-intensive bets at once. Based on the company’s disclosures and what management is emphasizing, here are the risks I’m watching into 2026:

• Execution risk in AI infrastructure: if AWS mis-times capacity, can’t secure power, or fails to deliver Trainium3 in volume, growth could lag peers even if demand is strong.

• Competitive intensity across retail, grocery, cloud, and ads is real and highlighted in Amazon’s own risk disclosures.

• Margin volatility: depreciation from data center build-out can pressure AWS margins and GAAP profits near term.

Final Thoughts

The real Amazon debate is whether Amazon is entering a phase where automation and AI convert scale into structurally higher margins in commerce, while AWS re-accelerates on a foundation of capacity and custom silicon. The evidence trail is already visible: in commerce with 1 million robots deployed with an AI control layer to raise throughput and lower unit costs; same-day perishables and rural logistics expansion to increase frequency and Prime value into 2026. In AWS, it’s capacity expansion plus a blended compute stack - NVIDIA where it’s best, and proprietary silicon where cost-performance and supply control become strategic. Layered on top is another profit pool: advertising and Prime Video, where incremental yield can translate into outsized profit dollars because shopping intent and attention are already there.

This is why “Amazon spends too much” is usually the wrong critique. The question isn’t whether the CapEx number is big; it’s whether the returns show up in the scoreboard. In commerce, I’m watching for falling cost-to-serve reflected in a higher (not perfectly linear) North America operating margin, plus evidence that perishables and rural expansion increase frequency without degrading unit economics. In AWS, I’m watching for capacity availability translating into sustained growth and durable operating income dollars, along with increased adoption of Trainium/Graviton chips. In ads and Prime Video, I’m watching for sustained ad revenue growth and improving profitability driven by expanding inventory (especially live sports) and broader DSP reach.

If Amazon executes across those three lanes at the same time, the next years should look less like spending and more like a compounding step-up in earnings power and cash flow.

His claim: Alphabet (), Amazon.com (), Meta (), Microsoft (), Oracle () are simultaneously ramping AI CapEx and stretching the “useful life” of their NVIDIA-powered data center hardware. In his numbers, that could understate depreciation by roughly $176 billion between 2026 and 2028 and leave reported operating income at companies like Oracle and Meta more than 20% above what he sees as economic reality.

Burry has backed this view with put options on NVIDIA () and Palantir (), the poster children of the AI boom, and doubled down after NVIDIA’s Q3 2025 earnings, arguing that investors are confusing “physical utilization” of older GPUs with true economic useful life. Rather than shorting the hyperscalers directly, he’s made a capped-downside, leveraged bet1 against the most sentiment-driven beneficiaries of the AI boom: the chip vendor at the center of the hyperscalers’ CapEx mega-cycle and a highly-valued pure-play AI software name.

If the economics or depreciation assumptions crack, hyperscalers can slow CapEx and watch free cash flow surge. NVIDIA and Palantir can’t: NVIDIA’s growth is tied to GPU budgets, and Palantir trades at a sky-high 223x TTM P/FCF. Burry’s trade is a focused bet against AI euphoria, not against the diversified cash machines like Alphabet, Amazon and Microsoft.

If you own those hyperscalers, as I do, you can’t dismiss this as cranky tweeting. You need to understand the accounting, what’s a genuine concern, and what’s mostly optics.

GAAP Depreciation, Accrual vs Cash, and Where CapEx Really Bites

To understand Burry’s accusation, we need to revisit the difference between cash and accrual accounting in the context of AI chips and servers. In earlier articles I walked through how depreciation and other non-cash items flow through the financial statements.

Suppose Google spends $900,000 on GPU racks this year:

• Cash reality (cash flow statement): the full $900k leaves the bank today and shows up as CapEx under Investments In Property, Plant & Equipment.

• Accrual / GAAP reality (income statement): you don’t expense $900k immediately. Management estimates a useful life, say 3 years, and spreads the cost evenly: $300k of depreciation each year as Cost of Goods Sold (COGS) or Selling, General, and Administrative (SG&A).

Economically, it’s one big cash hit upfront. In accounting terms, that hit is smoothed over several years as depreciation, so operating income (EBIT) shows a steady annual expense rather than a single lump. On the balance sheet, the GPUs sit in Property, Plant & Equipment (PP&E) and are gradually reduced by accumulated depreciation.

Depreciation is a non-cash charge: it lowers operating income and PP&E, but gets added back in operating cash flow because the cash already left when the CapEx was made. That’s why:

Operating Cash Flow (OCF) = Net Income + Depreciation (and other non-cash charges) ± Working Capital

Free Cash Flow (FCF) = OCF - CapEx

Now tweak just one assumption, useful life, and the optics change dramatically. Using the same $900k:

• If you depreciate over 3 years → $300k per year hits the income statement.

• If you depreciate over 6 years → only $150k per year hits earnings.

Same cash out, same hardware, but under a 6-year schedule, GAAP operating income looks much higher in the early years. That tension is exactly what Burry is attacking.

What Hyperscalers Have Actually Done With Depreciation

Across the industry, server lives have drifted out from roughly 3-4 years to 5-6 years just as AI CapEx has exploded.

In early 2023, Alphabet extended the useful life of its servers from 4 to 6 years, and certain network equipment from 5 to 6 years, cutting its 2023 depreciation by $3.9 billion and boosting earnings by $3 billion. Its investor FAQ adds that depreciation lives are “regularly evaluate[d] for factors such as technological obsolescence and our planned use and utilization”, reinforcing the idea that this is meant to track economics, not just massage optics.

Amazon has twice lengthened, and then partially shortened, the lives of AWS servers and networking kit. A Q4 2022 change from 4 to 5 years for servers (and 5 to 6 for networking) reduced 2022 depreciation by $3.6 billion, and lowered net loss. In Q4 2023, Amazon went from 5 to 6 years for servers, adding an expected $3.1 billion to 2024 operating income. Then, in Q1 2025, it shortened the life of a subset of servers from 6 back to 5 years, citing the “increased pace of technology development, particularly in the area of artificial intelligence and machine learning”.

Meta has moved from 3 years in 2020 to 5.5 years by 2025 through a sequence of extensions.

The pattern is clear: an industry that once assumed 3-4-year refresh cycles now depreciates much of its fleet over 5-6 years. That matters when the numbers are this big: Alphabet is expected to spend over $90 billion on CapEx in 2025 and Amazon more than $125 billion, with most of it going into AI infrastructure. Under 5-6-year schedules, only a slice of that shows up as depreciation each year.

Share

Burry’s Case: The CapEx Hyper-Cycle and Earnings Mirage

Burry’s argument is that AI hardware really lives 2-3 years, but is being treated as if it lives 5-6. NVIDIA’s flagship GPUs are on a 3-year cycle, with each generation 2-3x more efficient per unit of compute and per watt, so keeping old chips in service quickly becomes uneconomic for state-of-the-art training. Yet just as hyperscalers ramp CapEx to unprecedented levels, they’ve lengthened depreciation schedules to 5-6 years, which mechanically flatters earnings. Burry estimates that from 2026-2028, depreciation will be understated by about $176 billion, causing hyperscalers to overstate profits by more than 20%. If the true economic life is closer to 3 years, a cliff eventually arrives: either via large impairments when “stranded” chips are written down, or via the realization that what is considered growth CapEx is actually permanent sustaining CapEx, locking these companies into structurally lower free cash flow.

Bloomberg, summarizing this concern, has warned that hyperscalers could collectively spend trillions of dollars on AI infrastructure over this decade, with the risk that hardware lifespans prove shorter than the 5-6 years management currently assumes.

In that scenario, earnings quality deteriorates and valuations reset: optimistic useful lives + explosive CapEx + high multiples = vulnerable earnings and fragile sentiment.

The Bull Case: Longer-Lived Chips, Smarter Data Centers, and “CapEx Now, FCF Later”

The bullish view starts from a different premise: AI hardware and data center economics really have improved, and the accounting is trying to catch up rather than inventing a story.

Hyperscalers and NVIDIA argue that GPUs don’t drop dead after three years. The typical pattern is:

• Years 1-3: frontier-model training on the latest chips.

• Later years: those same GPUs are repurposed for inference (serving already-trained models to users), smaller models and internal workloads.

During NVIDIA’s Q3 FY2026 earnings call, CFO Colette Kress said that “thanks to CUDA [the software stack], the A100 GPUs we shipped 6 years ago are still running at full utilization today, powered by vastly improved software stack”, and emphasized that “GPU installed base, both new and previous generations, including Blackwell, Hopper and Ampere is fully utilized”. At an a16z infrastructure event in late October 2025, Google’s VP and GM of AI & Infrastructure, Amin Vahdat, said their 7-8-year-old TPUs are still running at 100% utilization. Peak use may be 2-3 years, but economic use can be much longer.

Engineering also helps the case for longer lives: modern data centers are modular (you swap GPUs and networking rather than ripping out racks), cooling and power management are better, and performance gains per generation are moderating, with more of the improvement coming from software and systems optimization. During Microsoft’s Q1 FY2026 earnings call, CFO Amy Hood has stressed that assumed lifetimes on some “short-lived assets” are matched to multi-year cloud and AI contracts - the revenue is designed to cover the hardware over its GAAP life.

On this view, today’s “CapEx now, free cash flow later” story is credible. Operating cash flows at Alphabet, Amazon, Microsoft and Meta keep climbing, while management is reinvesting heavily in AI infrastructure because they see more demand than they can meet at attractive returns. Even if FCF margins settle lower than in the ultra-lean 2010s, absolute FCF can still be much larger on a far bigger revenue base. And, crucially, if returns disappoint, hyperscalers can pull the CapEx lever; the real pain would then fall on the supply side - GPU vendors and leveraged data center operators whose growth depends on the spending hydrant staying on full blast.

Burry’s Rebuttal: Utilization ≠ Value Creation

Burry’s response, which I think is intellectually fair, is that just because GPUs are busy doesn’t mean they’re economically productive. Older chips can be far less power-efficient, so once you factor in electricity, cooling and the opportunity cost of not deploying newer hardware, they may only be marginally profitable. GAAP “useful life”, he argues, is supposed to reflect the period over which an asset generates real economic benefits, not simply the time until it physically stops working or can be assigned some low-value task.

In that sense, Burry and NVIDIA’s CFO are almost talking past each other: NVIDIA is saying, in effect, “our chips don’t become landfill after three years; customers keep them busy”, while Burry’s question is, “yes, but are they earning enough incremental return, net of power and opportunity cost, to justify spreading their cost over six years?”. Both perspectives matter if we care about the true economics behind today’s AI earnings.

In-House AI Chips: Eroding (Slowly) the NVIDIA Dependency

Parallel to the NVIDIA GPU boom is a quieter story: the rise of in-house AI chips at the hyperscalers. GPUs remain the “Swiss army knife” of AI, but Google, Amazon and Microsoft are building custom ASICs (Application-Specific Integrated Circuits) to handle an increasing share of their AI workloads.

Google led with TPUs (Tensor Processing Units) in 2015, and is now on its seventh generation (Ironwood). Amazon followed with Inferentia (for inference) and Trainium (for training), now in its third generation; Anthropic is training on half a million Trainium2 chips in an AWS data center with no NVIDIA GPUs. AWS says Trainium delivers 30-40% better price-performance than “other hardware vendors” on its platform. Microsoft is pursuing a similar path with its Maia accelerators for Azure.

The logic is straightforward: custom ASICs are expensive to design, but at hyperscale they can lower unit costs, improve energy efficiency and give more control over the AI stack. NVIDIA GPUs remain the flexible, high-end workhorses; TPUs, Trainium and Maia are more like specialized power tools. Over time, a larger share of AI training and inference can run on these in-house chips, reducing the need to buy NVIDIA for every incremental workload and putting at least a partial lid on the AI CapEx bill.

Share Level-Headed Investing

How This Shows Up In the Numbers: Earnings vs. Free Cash Flow

You can see this tension clearly in the financials of Alphabet and Amazon.

Since the AI boom in late 2022, Google’s operating cash flow has climbed strongly at 17.7% CAGR to $151 billion on a trailing basis in Q3 2025, reflecting the underlying profitability of Search, YouTube, Cloud, etc. but free cash flow has grown slowly at 5.6% CAGR, because CapEx has exploded for AI data centers at 37% CAGR. While Google’s OCF margin has expanded from 33% in 2022 to 39% in late 2025, FCF margin has shrunk from 22% to 19%, as the company ploughs more of its operating cash back into infrastructure.

On an earnings basis (which benefits from extended depreciation), Alphabet trades at a low-30s multiple of net income. On a free cash flow basis (which reflects the full CapEx burden), the implied multiple is in the low-50s.

A similar pattern exists at Amazon. Over the same 3-year period, Amazon’s operating cash flow has surged at a 49% CAGR to $131 billion on a trailing basis in Q3 2025, and its OCF margin has expanded from 8% in 2022 to 19% in Q3 2025. Yet free cash flow sits at only $11 billion, implying an FCF margin of roughly 1.5% and a valuation around 230x FCF, even though its P/E in the low-30s doesn’t look extreme for a dominant aggregator.

Some of this simply reflects aggressive growth investment where management believes returns are high, but Burry’s point is that the widening gap between net income and cash is being flattered by depreciation assumptions.

Are Useful Lives Too Long?

My view is that it’s a judgment call, not outright cooked books (so far): Extending server lives from 3 to 4 years as data center design improves and workloads diversify is plausible. Stretching all the way to 5-6 years for rapidly advancing AI accelerators is more debatable, especially given effective economic life shrinking for leading-edge GPUs as upgrades accelerate. However, the changes have been fully disclosed, and so far there is no hard evidence (whistleblowers, enforcement actions, restatements) that this is a deliberate deception rather than aggressive optimism.

The bigger issue is structural CapEx intensity: Even if the useful life assumptions prove roughly right, AI is turning hyperscalers into more capital-intensive businesses than they were in the 2010s. Sustaining CapEx (the amount needed to maintain current performance) is likely to be higher going forward, leaving structurally lower FCF margins than the peak years of the ad/cloud/commerce boom.

If Burry is directionally right, the pain hits vendors and the marginal players first. If hyperscalers decide their returns on AI CapEx are lower than expected, they retain flexibility to moderate CapEx.

What This Means for Alphabet & Amazon in My Portfolio

For my own positions in Alphabet and Amazon, the accounting lives matter less than the basic question of cash in vs cash out: are these incremental AI dollars actually driving faster revenue growth (Search, Cloud, Ads, Prime and so on), widening moats through better products and stickier platforms, and growing operating cash flow faster than CapEx over a full cycle?

If the hyperscalers’ CapEx mega-cycle brings in strong 20-30% returns on invested capital (ROIC), rising operating cash flows and a credible path to stabilizing free cash flow margins in the 20s, I’m prepared to view them as more capital-intensive compounding machines than they used to be. That requires separating earnings optics from economic reality: As discussed earlier, Alphabet and Amazon look much cheaper on EPS than on FCF, so I anchor on OCF growth and CapEx discipline, not headline earnings.

Over the next few years, I’ll be watching three things very closely:

1. Useful lives vs reality: do Alphabet and Amazon start taking impairments on AI hardware or shorten useful lives again, especially for accelerators?

2. CapEx vs cash generation: does CapEx growth eventually slow while FCF accelerates, or do we end up on a permanent AI CapEx treadmill?

3. External pressure: do regulators or auditors begin publicly questioning useful life assumptions?

If those signals turn negative while the narrative stays euphoric, that’s my cue to reassess.

Final Thoughts

Burry is not always right, but he is almost never trivial. His AI short is not a lazy “bubble” call; it’s a very specific claim that earnings at the core of the AI boom are being flattered by optimistic depreciation and an AI CapEx spree that may prove more sustaining than growth. At a minimum, that should force any level-headed investor in Alphabet, Amazon or Microsoft to look past the headline EPS and ask: what is the real economic life of this hardware, and what does that imply for future free cash flow?

At the same time, the hyperscalers are not blind. They are running formal life-studies, adjusting assumptions (including shortening useful lives again when AI advances faster than expected), investing heavily in their own chips, backed by management commentary that AI and cloud demand remains very strong. The cash flow statements show robust, growing operating cash flows and management teams that have already earned the benefit of the doubt on big, long-duration bets.

Where I land today is simple: I take Burry’s warning seriously, but I don’t view it as a reason to abandon high-quality compounding machines whose economics I still like. It’s a reminder to anchor on cash flows and returns on incremental CapEx. If AI CapEx delivers high, durable returns, today’s accounting noise will fade into the background; if it doesn’t, the cracks will show up first in cash, not in carefully manicured income statements.

As an investor, my task is to watch the numbers, monitor what management does more than what it says, and be ready to change my mind if the cash stops showing up.

That article followed my first direct, stock-specific bet on China: initiating a position in PDD Holdings Inc. (, the parent company of Pinduoduo and Temu) in late August 2024 after its Q2 results and a sharp sell-off. Since then, a lot has happened in Chinese markets, and in my own thinking. Fast-forward to early November 2025: I’ve exited my PDD position at an average profit of about 27%, and I’ve redeployed the capital into two higher-conviction compounders: Uber Technologies () and Constellation Software (CSU.TO).

This article is about why I made that shift, and why it doesn’t mean I’ve changed my mind about China.

One Year On: My China Thesis Still Stands

The core argument of my China piece was simple. China is a hyper-competitive, intensely dynamic market where creative destruction is the norm, in contrast to the West’s more entrenched platforms. Its regulatory regime is active, sometimes abrupt, and often driven by social and long-term economic objectives, even if that creates short-term uncertainty. Combined with deeply negative sentiment, this has left many high-quality Chinese businesses trading at significantly lower valuations than Western peers. The risks are real, and so are the rewards for level-headed investors willing to do the work, accept volatility, and think in decades rather than quarters. The question was never whether the risks exist; they clearly do. The real question is whether the potential rewards justify them. I still believe the answer can be yes.

So why sell PDD? Because a country-level thesis is not the same as company-level conviction. It’s consistent to be constructive on China as an economy and yet decide that a specific company no longer fits the way I want to express that view.

The Original PDD Holdings Thesis, and What Went Right

When I initiated my position in PDD in late August 2024, several elements lined up. The company had just reported Q2 2024 results. The market reaction was harsh: investors focused on near-term commentary and guidance, and the stock sold off by 40% from recent highs. At the same time, sentiment toward Chinese equities in general was deeply negative. That combination, a fundamentally strong business hit by a sharp, sentiment-driven sell-off, is exactly the sort of setup I look for as a value investor, and I quickly built PDD into a 10% position.

My thesis was straightforward: PDD was a structurally advantaged e-commerce platform, with deep penetration in rural and lower-tier markets in China and an innovative approach to social commerce and C2M (Consumer-to-Manufacturer) globally. I believed I was buying a great business at a substantial discount to its intrinsic value.

Price-wise, things unfolded reasonably well. After my initial purchase, the stock rallied strongly as sentiment improved, helped by government stimulus measures that lifted many domestic equities. Over my holding period, I ultimately exited at an average gain of around 27%.

Superficially, that’s a Graham-style textbook “successful” investment: buy when pessimism is high, hold as the business continues to execute, and sell after the market re-rates the stock upwards. Success on paper doesn’t automatically mean a position deserves to remain in the portfolio. Over time, my discomfort grew, not with the business performance per se, but with what I didn’t know.

Why I Chose to Exit PDD Holdings

The decision to sell was not based on a view that PDD’s business is broken or that its long-term prospects are poor. I expect the company to do well over time. Instead, the choice came down to three intertwined considerations: disclosure quality, portfolio construction, and relative conviction.

Disclosure Quality and Management Communication

As my holding period lengthened, I grew increasingly uneasy about how little I truly knew about the underlying drivers of PDD’s performance. On paper, the reporting is compliant; in practice, the disclosures are opaque, with limited granularity on segments and unit economics, making it hard for a long-term, fundamentals-focused investor to build a clear mental model of the business.

That was reinforced by management’s communication style. On earnings calls, direct questions often received high-level answers that largely repeated strategy and priorities rather than adding much beyond the press release and slide deck. I don’t expect management to reveal their playbook, but I do expect enough transparency to see how they measure success and what trade-offs they’re making.

With PDD, I increasingly felt I was flying with part of the instrument panel covered, acceptable for a small “toe-in-the-water” position, but not for a larger holding. By contrast, Uber’s financial statements, segment reporting, and commentary make it much easier to reconcile the story with the numbers. I don’t need perfection; I need enough clarity to underwrite a position with confidence, and over time PDD drifted below that bar.

Commerce Exposure and Portfolio Construction

The second factor was portfolio construction. PDD sits in a space where I already own two leading businesses I consider higher-conviction: Amazon.com () and MercadoLibre ().

Amazon is a global leader in commerce, cloud infrastructure, and digital ads, with a long track record of execution and transparent reporting around its key segments. MercadoLibre is Latin America’s dominant commerce and fintech platform, with strong network effects and a payments ecosystem that has become central to its region’s digital economy.

In different ways, all three businesses - Amazon, MercadoLibre, and PDD - operate as commerce aggregators. They differ geographically and strategically, but at the portfolio level they contribute to a similar thematic exposure: the economics of online platforms connecting buyers and sellers, a topic I explored in more depth in a prior article, Undervalued by Design: Amazon, Alphabet & the Economics of Aggregation.

Over time, I realized my allocation to this theme, and its risks, was higher than I wanted. If my portfolio is to remain concentrated in my very best ideas, I cannot justify multiple sizable positions that rhyme economically, especially when my conviction level differs meaningfully between them.

In that hierarchy of conviction, Amazon and MercadoLibre sit at the top. PDD, despite its impressive growth and innovation, slipped to a clear third.

Opportunity Cost and Relative Conviction

The final piece of the thesis is opportunity cost. By mid-2025, it became harder to justify keeping a position in PDD when I had structurally higher conviction in two other ideas: Uber and Constellation Software. Uber, which I’ve written about separately in Uber Technologies Inc. (UBER) Q3 2025 Results: The Global Royalty on Movement Shows Its Earnings Power, is evolving into an asset-light “royalty” on real-world movement and local commerce, and I find its valuation appealing relative to its cash-generation potential. In late October 2025, I also began building a position in Constellation Software during what is currently its largest draw-down, with the stock down 30%+ from its recent all-time high, an entry point I consider attractive for a proven compounder of this quality. From a portfolio construction perspective, Uber and Constellation help diversify risk across business models and end markets, and both are supported by disclosure practices that align with my analytical process. Faced with the choice between holding PDD, where my relative conviction had diminished, and reallocating into Uber and Constellation, where my conviction and perceived risk-adjusted returns were higher, the rational move was to redeploy the capital.

What This Move Doesn’t Mean About China

Exiting PDD does not represent a U-turn on my China thesis. It doesn’t mean I now think China is “uninvestable”. It doesn’t mean I’ve concluded Chinese companies cannot be good long-term holdings.

I expect that I will once again own Chinese companies in the future, potentially at greater size. When that happens, I want my capital in businesses where I can underwrite the long-term economics and governance with more confidence than I eventually could with PDD.

Process Takeaways: Conviction, Transparency, and Opportunity Cost

The PDD journey, from initial buy in August 2024 to exit in early November 2025, leaves me with a few process reminders that are worth writing down.

First, a macro thesis (“China will remain a critical growth engine for the world”) is not enough. For me, stock selection happens at the micro level: the specific business model, the quality of management, the governance structure, and the clarity of the information I receive as a minority shareholder.

Second, transparency acts as a kind of “meta-moat” around conviction. Two companies with similar underlying economics can feel very different to own if one communicates clearly and the other remains opaque. The stronger the business, the more painful it can be to admit that I don’t have the visibility I need. That admission is an essential part of disciplined investing.

Third, concentration is meaningful only if it is reserved for the ideas where my conviction is highest, not the businesses that fit a theme. The more I refine my process, the more willing I am to ask, “If I started with fresh capital today, would I still choose this stock over the alternatives?” If the answer is no, that’s a signal I need to pay attention to, even if the position has done well.

Finally, opportunity cost is not an academic concept. Every dollar tied up in PDD was a dollar that couldn’t be working for me in Uber, Constellation Software, or another high-conviction idea. Taking a reasonable profit and rotating into better-aligned opportunities is part of long-term thinking.

Final Thoughts

I continue to view China as one of the most important and fascinating arenas for level-headed investors, a place where risk and opportunity are tightly intertwined, and where understanding the local context is crucial. My October 2024 article on investing in China reflects my view of that landscape at a high level.

At the same time, my decision to exit PDD and redeploy capital into Uber and Constellation Software shows how I’m refining the way I choose to invest my capital: I’m being more demanding about the combination of business quality, governance, disclosure, and fit with my overall concentration.

My portfolio is not a museum of past theses; it is a living, evolving expression of where I believe risk and reward are most attractively aligned today. Right now, that means not owning PDD, and owning Uber and Constellation Software.

What Q3 Results Said

Revenue fell 2.4% YoY to €507.1 million - Evolution’s first YoY decline and the 12th sequential quarter of growth deceleration.

Live Casino revenue (85% of mix) contracted 3.4% YoY; RNG grew 4.1% YoY (15% of revenue).

During the Q3 earnings call, CEO Martin Carlesund, commented on the revenue decline, saying:

Regionally, Asia broke. Revenue slid to €189 million (-6.5% YoY, -9.6% QoQ) as counter-piracy measures overshot and hurt legitimate users. Management’s tone shifted from cautious optimism to acknowledging they over-extended actions and still don’t have a time frame for normalisation. Europe finally ticked up sequentially after the full ring-fencing effect, but remained -6.5% YoY at €182 million. North America grew +14.5% YoY to €74.2 million, Latin America +6.4% YoY to €39.8 million.

Operating leverage stalled. OpEx grew +5.3% YoY but fell -3.4% QoQ as the company pivoted to cost control. Headcount declined sequentially (21,652 vs. 22,248 in Q2) even as it was still +4.2% YoY; Game Rounds Index slipped -0.6% YoY and -8.03% QoQ.

Cash flow stayed strong. Q3 operating cash flow was €371.9 million; with €29.8 million of CapEx (intangible + tangible), free cash flow was €342.1 million (a modest 4.2% YoY growth) and cash conversion 83%.

Regulated mix keeps rising - 46% in Q3 vs. 45% in Q2. Consistent with the strategic pivot I highlighted in Q1, but it is not yet translating into growth.

Legal overhang twists the narrative. After a New Jersey court saga, Playtech was identified as the client behind the 2021 smear dossier (via Black Cube); Evolution will expand its defamation case; Playtech contests the allegation. This may prove financially immaterial but it underscores the intensity of competitive dynamics.

Where Q3 Hits My Q1 Thesis

Back in May I laid out a “friction now, re-acceleration later” case; Q3 tested each pillar of my investment thesis:

1. Europe’s ring-fencing would reset the base and allow sequential growth.

   This part is playing out, Europe returned to +1.1% QoQ growth, but QoQ stabilization isn’t offsetting Asia’s damage.

2. Asia’s cyber issues were a solvable execution problem with long-term upside intact. Management now admits they over-did the countermeasures and still lack a resolution. After over a year of firefighting, this looks structural, not episodic.

3. Capacity additions: Studios in Brazil, the Philippines; dedicated North American environments would monetize in H2.

   New sites are off to a good start, and North America/LatAm grew, but the Game Rounds Index declined and Live Casino revenue fell; capacity isn’t yet translating into system-wide activity.

4. Profitability resiliency would bridge the growth gap.

   EBITDA margins are still within the 66-68% guided range, and cash generation is stable, but operating leverage has turned negative for three quarters, forcing a sharper cost/resource mix optimization focus.

Put simply: Q3 converts a “temporary slowdown” into a handicap: Asia’s drag appears persistent, not transitory, and the company is managing for stability rather than market share gains.

Is Asia’s Piracy Problem Fixable, or Just Manageable?

Management says they’re spending “whatever is required” , yet also admits they overshot and hurt legitimate users, suggesting a tuning problem rather than a solvable one-off fix. As Carlesund put it:

When asked whether more spending could accelerate a fix, he was blunt:

To me, that confirms a cat-and-mouse dynamic: not a budget constraint, but a structural reality that must be continually managed via trade-offs; tighten too much and you block real users; loosen and pirates slip through. Carlesund also offered no timeline:

One important nuance: the defensive work could become an edge beyond Asia:

If piracy in semi-regulated pockets is an enduring tax on growth, Evolution likely needs to rethink go-to-market in those markets, beyond brute-force technical countermeasures, while proving that its strengthened core really does widen the moat globally. I can’t yet see that in the numbers.

India’s 2025 Online Gaming Law

India has tightened the backdrop for Evolution’s B2B supply without banning all online gaming: Parliament passed the Promotion and Regulation of Online Gaming Act, 2025 in August. The Act prohibits “online money games” (real-money stakes). Carlesund noted on the Q3 call:

Despite Carlesund comment, to me, this functions as a de-facto total ban on real-money online casino in region, impacting operators using Evolution’s B2B solutions to move money, and it implies a clear financial headwind for Evolution.

Share

The Fixed-Cost Reality

Evolution’s model carries a large fixed-cost base, notably 24/7 human presenters. When activity (game rounds) softens or geographies are gated, revenue per Full Time Employee (FTE) falls and operating leverage turns against the company. Q3 shows sequential OpEx cuts and a sequential headcount decline, a conspicuous shift from “hire to grow” toward “defend margins while we troubleshoot demand”. That’s prudent, but emblematic of a more volatile phase of the business.

Historically, Evolution only trimmed its workforce in 2020 (COVID studio measures) and 2024 (the Georgia strike), not due to weak demand. By contrast, Q3 marks the first deliberate, demand-driven downsizing, aligning with management’s own principle that “we don’t hire unless we grow”. Read plainly, this is a regime shift for a fixed-cost model: from scaling capacity on a rising tide to protecting margins amid low-to-negative growth expectations.

CapEx discipline is similar: with €29.8 million in Q3 and management signalling full-year CapEx “likely on the high side” versus the €140 million plan, the investment cadence is easing. This does not resemble the posture of a company seeing unconstrained demand and on the offensive to gain market share.

Bright Spots, And Why They’re Not Enough Yet

Europe’s base looks reset. After the painful cleanup, sequential growth suggests the ring-fencing phase is largely behind them. The UKGC licence review is still pending, but management reported no new developments in Q3.

North America & LatAm are working. Dedicated studios are going live across the U.S./Canada; Brazil is scaling ahead of plan. These regions can grow for years, but the mix is lower-margin than Asia, so earnings growth will lag revenue unless Asia stabilizes and returns to growth.

Cash remains a fortress. With €656.4 million in cash after €187 million of Q3 buybacks (YTD €406.5 million), Evolution can keep rewarding shareholders by continuing to buy back shares at low prices. The margin guide stayed intact.

Valuation Is No Longer The Whole Story

Q3 delivered €342.1 million of free cash flow. The balance sheet remains pristine and buybacks are material. However, the market is not merely discounting a few soft quarters; it is questioning the trajectory. The revenue decline, coupled with a more cautious CapEx cadence, and explicit cost discipline, signal business uncertainty rather than re-acceleration. Until Evolution demonstrates a credible path to normalizing Asia, re-accelerating growth, and restoring operating leverage, the stock price’s multiple will likely embed that uncertainty: Evolution stock now trades at 10.5x Price-to-Free-Cash-Flow (TTM), near its all-time low multiple.

Share Level-Headed Investing

Revised Positioning: Hold With a Bias to Trim

I rated Evolution a hold in Q1 at an undemanding multiple, expecting improvement by late-2025. Post-Q3, I downgrade to hold/trim given the increasingly structural Asia risk, negative operating leverage, and a managerial pivot toward defense (cost control, tighter CapEx). My position has drifted from 10% (in late 2024) to 6% purely on stock price decline; I have not added since February 2025 and do not intend to buy the dip. I’ll keep Evolution as is and redeploy any trims into higher-conviction ideas, with new capital contingent on tangible progress against the following triggers:

• Asia stabilization: Consecutive quarters of positive QoQ revenue and Game Rounds Index growth without new countermeasure shocks.

• Europe momentum post ring-fencing: Continued sequential growth with regulated-mix gains not cannibalizing the top line.

• Operating leverage: Revenue/FTE and free cash flow margin trending up alongside activity, not just via OpEx cuts.

• North America monetization: Sustained double-digit growth with clearer throughput from new dedicated studios and meaningful contribution from Galaxy Gaming acquisition.

• CapEx cadence: A re-acceleration from today’s cautious posture, signaling demand visibility.

• Regulatory overhangs: Satisfactory conclusion of the UKGC license review and no new ring-fencing waves.

Final Thoughts

After Q3 2025, the “temporary slowdown then re-acceleration” view no longer holds; confidence now depends on demonstrated improvement.

The business is intact: market leadership, fortress cash generation, and enviable margins, but the profit engine (Asia) is impaired by an adversary that money alone cannot solve. Management’s own words now emphasize volatility, caution, and cost discipline rather than market share-taking growth. Europe’s base is reset and the Americas are progressing, yet the mix shift is lower-margin and operating leverage is negative. Until Evolution demonstrates Asia stabilization, renewed activity (game rounds), and the return of operating leverage, the multiple will likely remain anchored near trough levels despite stable free cash flow.

Not royalty in a legal sense, but economically: a structurally advantaged platform that clips a small fee on a rising volume of real-world transactions (rides, food, groceries, retail, on-demand logistics) without owning the underlying physical assets. I borrow this framing from Bill Ackman, who described the ideal business as one that “collects a royalty on the growth of others” in an interview that shaped my thinking about platforms.

Q3 shows this idea playing out in practice. The business is compounding scale and engagement not by pushing price, but by making the platform more useful, more integrated, and more affordable. CEO Dara Khosrowshahi captured this shift in tone and posture in the earnings press release:

And he made the strategic direction explicit:

Underpinning this “royalty” profile is a liquid, defensible marketplace: a broad, global base of earners; dense geographic coverage; and a deeply integrated consumer experience across Mobility and Delivery. Cross-platform users are more loyal and more valuable, yet penetration of multi-product usage is still low enough to leave a long runway. This quarter’s acceleration is not just more usage; it is the network effect compounding: more users, using Uber more often, across more categories, with increasingly high switching costs for both consumers and merchants.

Five Key Takeaways From Q3 2025

The quarter combined powerful top-line momentum with more measured, but still positive, growth in profitability and cash generation.

Scale and Engagement

Total trips reached 3.5 billion (+22% YoY), Gross Bookings $49.7 billion (+21% YoY), and Monthly Active Platform Consumers (MAPCs) 189 million (+17% YoY), with monthly trips per MAPC up 4%. These figures aren’t an anomaly; they sit broadly in line with (or slightly below) Uber’s multi-year compounding: Total trips have grown at 24.3% over the past five years, Gross Bookings at 27.5%, and MAPCs at 19.5% CAGR. In other words, Q3 confirms that the long-term growth engine is intact; driven by real usage and frequency rather than price inflation, with average pricing roughly flat.

Revenue

Revenue grew to $13.5 billion (+20% YoY), broadly in line with Gross Bookings, reflecting healthy unit economics while the platform scales.

Operating Income

Income from operations (EBIT) increased to $1.1 billion (+4.7% YoY). In isolation that looks like steady, unspectacular growth; in context, it’s strong evidence of operating leverage. Since Uber first turned sustainably profitable at the operating line in mid-2023, operating income has compounded 72.4% yearly from $326 million to over $1 billion per quarter, growing faster than revenue over this period. That progression shows Uber adding incremental Gross Bookings and revenue without a proportional increase in Cost of Goods Sold (COGS) and operating expenses. In other words, a larger share of each new dollar flowing through the platform is dropping to the operating line.

Free Cash Flow

Trailing twelve-month (TTM) FCF reached $8.7 billion (+45.3% YoY). The cash engine remains solid, even if quarterly growth is muted by timing effects and reinvestment.

Network Quality and Cross-Platform Leverage

Uber reported 9.4 million drivers and couriers earning around $22 billion in Q3, reinforcing marketplace depth. Consumers using both Mobility and Delivery spend 3x more and have 35% higher retention, yet only 20% of eligible monthly active users currently use both products; an embedded, high-ROI growth lever. Khosrowshahi’s own summary on the call, “an outstanding quarter” with trips and Gross Bookings growing strongly while pricing stayed flat, aligns with this picture of durable, demand-led growth.

Subscribe now

Uber’s Real Profit Story Runs Through FCF

The headline GAAP number, Net Income of $6.6 billion, is eye-catching and, on its own, misleading. As disclosed in the press release and detailed in the 10-Q, this figure is heavily distorted by two non-operational items:

• A $4.9 billion benefit from the release of the Netherlands deferred tax assets valuation allowance which should reduce future GAAP tax expense / effective tax rate over time; and

• A $1.5 billion net unrealized gain from revaluations of equity investments (notably Aurora, Didi, and Grab).

The tax item is purely non-cash. Uber has concluded it is now “more-likely-than-not” that these deferred tax assets will be realizable, which justifies releasing the allowance. That is an economically positive milestone, but it inflates Q3 Net Income without reflecting incremental cash generation. CFO Prashanth Mahendra-Rajah is transparent about this, in the prepared remarks:

This is one of the reasons why I anchor on Free Cash Flow (FCF) as the measure of earnings power. In Q3, FCF was $2.2 billion, up 5.7% YoY. TTM FCF was $8.7 billion, up from $5.95 billion in the prior-year period, +45.3% YoY. Mahendra-Rajah reinforces how to read it:

The modest 5.7% FCF YoY growth in Q3 is not a red flag. It reflects normal working capital and payment timing effects, and a deliberate choice to keep investing in growth initiatives (Grocery & Retail, membership, merchant tools, autonomy) instead of maximizing near-term reported profits. On the call, the CFO underlined this philosophy, noting they are “very deliberately moderating the pace of our margin expansion” because they “have so many exciting opportunities to invest in”.

Stripped of one-time accounting noise, the picture is straightforward: Uber is sustainably profitable on a GAAP operating basis and strongly cash generative, with FCF growing healthily over a 12-month lens.

Mobility: The Core Royalty on Urban Movement

The Mobility1 segment is the clearest expression of Uber’s royalty economics.

Q3 2025 Mobility Gross Bookings reached $25.1 billion, up 19.6% YoY, while Mobility revenue rose to $7.7 billion, up 19.9% YoY. Mobility trips grew 21%, outpacing already-strong historical run-rates. Looked at over the past five years, the chart makes the story clearer: Mobility Gross Bookings have compounded at mid-30s%, segment revenue at low-40s%, and Mobility adjusted EBITDA2 at more than 50% CAGR annually over that period. As the five-year trend shows, that underlying profitability has expanded dramatically as the network has densified.

Khosrowshahi framed the step-change:

Several dynamics underpin this performance. Uber pushed affordability, including lower-cost products and moderating insurance pressures, without sacrificing profitability at the platform level. Average price remained roughly flat (–1% YoY) while trips and Gross Bookings surged. Geographically, growth was broad-based: strong U.S. volume on improved affordability and international strength supported by summer travel, taxi integrations, and formats like Reserve, Moto, and Uber for Business.

Liquidity is a key edge. An expanded and better-utilized driver base shortens wait times and increases reliability, reinforcing user preference and pushing more demand onto the platform. The 9.4 million drivers and roughly $22 billion in payouts in Q3 highlight the depth of that ecosystem.

For investors, Mobility demonstrates that Uber can maintain and extend category leadership while keeping its service affordable and still producing meaningful operating profit and cash flow. It is the anchor royalty stream on urban transportation.

Delivery: From Food to Full-Stack Local Commerce

If Mobility is the original royalty on movement, Delivery3 is the emerging royalty on local commerce.

In Q3 2025, Delivery Gross Bookings grew 25% to $23.3 billion, and Delivery revenue grew 29% to $4.5 billion. The gap between Gross Bookings growth and revenue growth is important: it reflects not only volume, but also better monetization. The five-year trend reinforces this: Delivery Gross Bookings have climbed from $8.6 billion to over $23 billion, while Delivery revenue has nearly quadrupled, compounding at about 31.5% vs. 22.2% for Gross Bookings. That spread, together with the shift from losses to solid positive adjusted EBITDA, is a clear sign of improving take rates and structurally stronger monetization.

The 29% revenue growth is attributable to the 25% Gross Booking increase plus a $146 million increase in advertising revenue, a high-margin layer on top of the logistics network. This is exactly how a platform royalty should evolve, by earning more per unit of activity without undermining user value.

The strategic centerpiece is Grocery & Retail (G&R). Management highlights that G&R is “an approximately $12 billion Gross Bookings run rate and growing significantly faster than restaurant delivery”. Q3 2025 results show that G&R drives customer acquisition: about 7% of Delivery first-time users come through G&R and platform frequency increases by 27% after users adopt G&R.

Meanwhile, Uber is becoming a growth engine for merchants: Over 1.2 million merchants use Uber, and ads plus merchant-funded offers are scaling quickly, with spend “up over 50%” and driving incremental sales.

This is Delivery maturing from a food-delivery marketplace into infrastructure for anything local - a second, diversified royalty stream on commerce, reinforced by data, membership (Uber One), product innovation (Meal Deals, quality improvements), and partnerships.

Freight: Cyclical, Contained, and Non-Core

The Freight4 segment remains the most challenged piece of the business, reflecting the broader freight downcycle rather than a structural break in Uber’s model. In Q3 2025, Freight Gross Bookings and revenue were $1.3 billion, flat YoY. Management attributes this to a “challenging freight market cycle”, with lower revenue per load offsetting volume, keeping Freight Gross Bookings “relatively flat”.

Segment losses accounted for a modest -$20 million in the quarter. Two points matter: First, Freight is not consuming disproportionate capital or management bandwidth relative to the scale of Mobility and Delivery. Second, if and when the freight cycle turns, Uber has the infrastructure to benefit without rebuilding from scratch.

For now, Freight is best treated as neutral-to-modest optionality, a “call option” on a cyclical recovery, not a pillar of the thesis.

Subscribe now

Capital Allocation & Balance Sheet: Quietly De-Risking While Buying Back Stock

Uber exits Q3 2025 with a balance sheet that is strong, liquid, and getting cleaner. The company holds about $9.1 billion in unrestricted cash, cash equivalents, and short-term investments, plus roughly $10.3 billion in equity stakes, against $11.8 billion of debt that is largely long-dated, fixed-rate Senior Notes. YTD, Uber has monetized approximately $1.4 billion of those equity stakes and is recycling the proceeds into growth and balance sheet optimization, not plugging holes.

In Q3, Uber issued $2.3 billion of new Senior Notes and used them to retire higher-coupon 2027 and 2028 notes and position itself to pay down the $1.2 billion Convertible Notes due December 2025; extending average maturities by over a year while keeping its weighted average cost of debt roughly flat, as management highlighted in the prepared remarks. Alongside this, Uber repurchased about $1.5 billion of stock in Q3 (roughly $4.6 billion YTD) and authorized an additional $20 billion buyback program, all funded out of robust cash flow. The stated hierarchy is clear and credible: reinvest in high-return initiatives, maintain ample liquidity and an investment-grade profile, and return genuinely excess cash to shareholders.

Looking Ahead: Scaling the Royalty Machine

Looking ahead, management’s commentary and guidance are consistent with the royalty narrative and the financial results: sustain high-quality growth, keep generating cash, and reinvest where the platform’s advantages compound.

On the medium term, management reiterates the framework first outlined during the Investor Update in February 2024 and echoed on the call: mid-to-high teens Gross Bookings growth over a multi-year period, while maintaining strong profitability and cash generation. Mahendra-Rajah stated:

The strategic levers for the next years are clear in Khosrowshahi’s remarks:

• Deepen trip → lifetime experience: push cross-platform engagement well beyond the current 20% of monthly active users using both Mobility and Delivery.

• Scale Grocery & Retail and “get anything nearby” into a much larger local commerce royalty stream of the $10+ trillion addressable market.

• Growth engine for merchants by expanding tools, integrations, and ads so restaurants and retailers rely on Uber for both demand and logistics.

• Build a hybrid human-autonomous marketplace, with Uber expecting to be live with autonomous vehicle deployments “in at least 10 cities by the end of 2026”, supported by its expanded NVIDIA partnership and marketplace data.

• GenAI everywhere. Use AI to price, match, route, and support more intelligently, lifting efficiency and experience without heavy incremental cost.

Valuation: Paying a Modest Multiple for a Growing Business

When I first wrote about Uber, my thesis was simple: once the business proved it could consistently generate free cash flow, the market would have to re-rate it from “story stock” to “cash machine”. The last two years have validated the first half of that view; the re-rating, in my opinion, is incomplete.

The free cash flow chart tells the story. Since mid-2022, TTM FCF has climbed from $672 million to $8.7 billion (119% CAGR), while the P/FCF multiple has compressed, not expanded. In other words, cash generation is up many-fold; the price investors are willing to pay for each dollar of that cash is lower than it was when the business was riskier and unproven.

On today’s numbers, the market is effectively valuing Uber at 22x TTM FCF. For an asset-light, scaled, global platform that is guiding to growing Gross Bookings in mid-to-high teens, already produces multi-billion dollar annual FCF, has visible reinvestment runways in G&R, merchant solutions, ads and autonomy, and operates with structural advantages in liquidity, coverage, data and cross-platform engagement, that multiple looks conservative. A low-teens P/FCF would imply deep skepticism that this royalty on movement and local commerce is durable; a low-20s multiple assumes it is, but barely credits the length or quality of the runway.

To frame it numerically (directionally, not precisely), I start with $8.66 billion of TTM FCF. Management guides to a “mid-to-high teens” Gross Bookings CAGR and is “targeting adjusted EBITDA growth at more than 2x Gross Bookings growth”, which implies roughly 30-40% growth in their profitability proxy. If Uber were to sustain its current 100%+ FCF conversion on adjusted EBITDA, that framework could support ~32-43% FCF CAGR.

I view that as possible, but optimistic. For my base case, I haircut this to a conservative 20% FCF CAGR over five years - my 20% FCF CAGR assumption sits below management’s framework of profits growing >2x Gross Bookings, leaving room for execution slippage. I assume diluted shares outstanding grow a modest ~1% over the same period (conservative, given buyback programs), and I hold the exit multiple at 22x P/FCF, roughly where the market has traded recently. On those inputs, year-5 FCF is $21.55 billion; at 22x P/FCF and on the projected share count, that implies a future share price a bit above $210. Discounted back at 10%, I get a current intrinsic value of $131 per share, which, against a current price of around $92, suggests Uber is trading at ~40% below my estimate of fair value. On these conservative assumptions, that gap implies the stock could compound at ~18% CAGR from today’s price, comfortably above my 10% hurdle rate.

This is not a precise price target, but a sanity check: under reasonable, deliberately conservative assumptions, today’s valuation leaves meaningful upside if Uber executes on its stated plan - compounding Gross Bookings, increasing cross-platform selling opportunities, and scaling profits faster than revenue.

In line with that view, I’ve been building my position in Uber since late May 2025, using a mix of new monthly contributions (dollar-cost averaging) and capital recycled from a lower-conviction holding I exited at a 27% profit in 14 months. Today, Uber sits at approximately 6% of my portfolio, with a path to build it toward ~10% over time to reflect my conviction in the business, its growth runway, and the current undervaluation, while still respecting the residual risks.

Final Thoughts

Uber is no longer a “someday story”; Q3 2025 confirms it’s already operating as a cash-generating, scaled infrastructure platform on top of real-world demand. The GAAP headline is noisy, but the underlying picture is clean: entrenched Mobility, a rapidly evolving Delivery business pushing into local commerce, clear operating leverage, disciplined reinvestment, and billions in free cash flow to show for it. On my conservative assumptions, 20% FCF CAGR, modest share creep, a 22x P/FCF exit, the stock screens roughly 40% undervalued, implying around 18% annualised returns from here, comfortably above my hurdle rate without needing heroic AV or blue-sky scenarios. Risks around regulation, competition, and execution are real, but at today’s price I believe I’m being more than fairly compensated to underwrite them. For me, Uber now sits where I want my capital: in a growing global royalty on movement and local commerce.

“This was a terrific quarter for Alphabet, driven by double-digit growth across every major part of our business. We are seeing AI now driving real business results across the company. We delivered our first-ever $100 billion quarter.”

— Sundar Pichai, CEO, Alphabet & Google

This quarter lands after a long stretch (2023 through early-2025) when headlines and investors doubted Google’s AI position, first after Bard’s botched debut wiped roughly $100 billion of stock market value in a day, then as Microsoft’s OpenAI bet seemed to widen the gap and Alphabet’s own CapEx plans stirred margin fears. I took the other side of that trade: I stayed bullish for years and kept adding on stock price’s weakness, arguing the stock didn’t reflect the fundamentals. I laid out that thesis in late-December 2024 - see my earlier post here.

For a flavor of the period’s skepticism, see the Bard error and selloff (Feb 2023), the cloud/AI comparison to Microsoft (Oct 2023), renewed margin fears as AI spend accelerated (Jul 2024), and the early-2025 sell-off tied to AI cost concerns.

The Full-Stack Advantage

During the Q3 earnings call, Pichai reiterated the company’s strategy succinctly:

On infrastructure, Google is scaling both NVIDIA GPUs and its own TPUs, “Our highly sought-after TPU portfolio is led by our 7th-generation TPU, Ironwood, which will be generally available soon”. On AI models and tooling, Gemini usage is exploding: “Gemini now processes 7 billion tokens per minute via direct API used by our customers. The Gemini app now has over 650 million monthly active users, and queries increased by 3x from Q2”. And distribution is broadening via products people already use - Search, YouTube, Android, and subscriptions, where Google has “crossed 300 million paid subscriptions led by growth in Google One and YouTube Premium”.

The research engine is still firing. Pichai highlighted a quantum breakthrough and teased the AI roadmap: “We are looking forward to the release of Gemini 3 later this year”. Meanwhile, token volume across Google surfaces has gone parabolic: from 980 trillion monthly tokens in July to over 1.3 quadrillion now, part of “more than 20x growth in a year”. Just as important as the sheer volume is the span of sources: those tokens arise from a mosaic of real-world interactions - search intent, YouTube viewing, Maps routing, Android usage, continuously coursing through Google’s infrastructure.

Services: Search, YouTube & Subscriptions Show Broad-Based Strength

Google Services revenue reached $87.1 billion, +14% YoY, an acceleration of 1% versus the prior year:

• Search & Other rose 15% to $56.6 billion, with retail and financial services the largest contributors and health care also adding to growth - and the drivers point to a healthy, expanding core, not erosion to competing AI chatbots.

• YouTube advertising increased 15% to $10.3 billion, driven primarily by direct response.

• Subscriptions, Platforms & Devices grew 21% to $12.9 billion on strong momentum in YouTube and Google One.

Services operating income was $33.5 billion (+9% YoY) with a 38.5% margin, both dampened by the European Commission (EC) fine; excluding the $3.5 billion EC charge booked in Services, operating income would have been $37 billion and operating margin 42.5%, a +2.2% expansion vs. 40.3% a year ago. These regulatory fines are episodic and largely one-off; for Google (and Big Tech generally) they’ve become a recurring cost of scale, more a periodic “tax on dominance” than a structural impairment to the business model.

The biggest strategic shift in Services is AI-native search - CBO Philipp Schindler captured the moment, “AI is driving an expansionary moment and transforming how people use Google Search”:

• AI Overviews have now scaled “to over 2 billion users”, and crucially, “overall we see the monetization at approximately the same rate”. The near-identical phrasing for two quarters implies ad load in AI responses hasn’t materially increased yet. Even so, sheer usage growth (billions of incremental queries at current monetization levels) is driving revenue while Google continues to refine ad formats and density.

• AI Mode has been rolled out globally across 40 languages and is gaining traction, with “over 75 million daily active users” in the U.S., “strong and consistent week-over-week growth”, and queries doubling over the quarter. In the 10-Q filing, Alphabet attributes the 15% YoY jump of Search to “increases in search queries resulting from growth in user adoption and usage on mobile devices; growth in advertiser spending; and improvements we have made in ad formats and delivery”.

Monetization moved in the right direction on both volume and pricing, with paid clicks +7% YoY and cost-per-click (CPC) +7% YoY, evidence that the ad machine continues to calibrate around these new surfaces. The data supports my view that fears of Search losing ground to AI chatbots were overstated.

YouTube remains a twin-engine model, ads plus subscriptions, inside a healthy creator flywheel. In the living room, YouTube has been #1 in U.S. streaming watch time for over two years, and Shorts in the U.S. now earn more revenue per watch hour than traditional in-stream.

On ad innovation, Demand Gen improvements drove 40%+ increases in conversion value for target-based bidding, and interactive direct-response ad formats in the living room have surpassed a $1 billion annual run rate. Subscriptions (Music, Premium, and YouTube TV) continue to rise, with Schindler noting that a paying user often “generates a meaningful higher gross profit” versus ad-supported alone.

Share

Cloud: Scale, Mix, and Margin

Google Cloud revenue grew 34% YoY to $15.2 billion. Enterprise AI products are now “generating billions in quarterly revenue” and pulling through both infrastructure (TPUs/GPUs) and solutions (Gemini 2.5 and peers). Operating income rose 84.6% YoY to $3.6 billion, lifting operating margin from 17.1% to 23.7% despite heavier depreciation and energy costs. More revealing, though, is what’s behind that margin expansion: Google’s structural cost advantage from TPUs, its custom AI chips. By leaning on in-house TPUs, Google sidesteps much of that “GPU tax” and enjoys a superior tokens-per-dollar1 profile. It’s not merely a cheaper chip, it’s a lower-cost stack end to end.

Demand is strengthening on three fronts. First, “the number of new GCP customers increased by nearly 34% year over year”. Second, deal size is inflecting: “We have signed more deals over $1 billion through Q3 this year than we did in the previous 2 years combined”. Third, adoption is deepening: “Over 70% of existing Google Cloud customers use our AI products”. AI monetization is scaling: “revenue from products built on our generative AI models grew more than 200% year-over-year” and nearly 150 customers processed ~1 trillion tokens each over the past 12 months.

Massive Backlog, Usage-Driven Revenue

Revenue backlog sits at $155 billion, up 46% sequentially and 82% YoY - roughly two and a half years of revenue at today’s run-rate. The 10-Q adds more color: remaining performance obligations (RPO) were $157.7 billion, “primarily related to Google Cloud”, with over 55% of that set to convert to revenue within 24 months. These figures reflect multi-year enterprise commitments from customers that view Google Cloud’s infrastructure as mission-critical. The filing also clarifies that RPO includes deferred revenue and amounts not yet invoiced but excludes cancellable arrangements and those with an original term of one year or less, and stresses that recognition depends both on Google’s ability to deliver and “when our customers utilize services”. In other words, the headline backlog is real demand, but the pace at which it turns into revenue is usage-driven.

On customer demand signals, Pichai highlighted a fresh Anthropic expansion: the AI lab “plans to access up to 1 million TPUs”. My take is that this deal is a powerful indicator of demand, but I’m cautious about assuming it fully materializes on the timetable or at the implied utilization, as the conversion risk sits with Anthropic’s economics - the company has been heavily loss-making, burning an estimated $5.6 billion in 2024 with an expected $3 billion burn in 2025, despite rapidly rising revenue. Accordingly, I discount this deal in my assessment of backlog, and my investment thesis does not rely on this AI lab deal materializing.

AI Inside: Real Productivity Gains

Alphabet isn’t just selling AI, it’s using it to run leaner and faster. On the call, Schindler said:

He added that in Customer Support, “Gemini-powered solutions have managed over 40 million customer sessions so far this year and resolved hundreds of thousands of customer inquiries”. CFO Anat Ashkenazi underscored the broader impact, noting that “now nearly half of all code [is] generated by AI” as part of ongoing efficiency work across the company. You can see the financial echo of those gains in management’s framing that margins benefited from “continued efficiencies in the expense base” (excluding the EC fine).

Outlook: Q4 2025 and 2026

Ashkenazi struck a constructive tone for Q4: “At the current spot rates, we could see an FX tailwind to our revenues in Q4. However, the volatility in exchange rates could affect the impact of FX on Q4 revenues”. Advertising comps will be tougher, “as year-over-year comparisons in advertising will be negatively impacted by the strong spend on U.S. elections in the fourth quarter of 2024, particularly on YouTube”. Cloud remains the engine, with “demand for our products” still high and a “$49 billion sequential increase in Cloud backlog in Q3”. To meet AI infrastructure needs, Alphabet “now expect[s] CapEx to be in the range of $91 billion to $93 billion in 2025”, and “we expect the growth rate in depreciation to accelerate slightly in Q4” She also noted that “sales and marketing expenses [will] be more heavily weighted to the end of the year”.

Looking to 2026, Ashkenazi guided to capacity-first execution: “[…] while we have been working hard to increase capacity and have improved the pace of server deployments and data center construction, we still expect to remain in a tight demand-supply environment in Q4 and 2026”, and “we expect a significant increase in CapEx” next year, with specifics on the next earnings call. The message is clear: near-term depreciation and operating cost pressure are the toll for extending the lead in cloud and AI, backed by visible demand. Depreciation rose 41% YoY in Q3 and, with CapEx roughly doubling, is poised to accelerate through 2026. Excluding the EC fine, operating margin came in at 33.9%, a shade below the 34%+ of the prior two quarters.

Valuation: Pricing the Full-Stack Compounding Engine

In my earlier article, Undervalued by Design: Amazon, Alphabet & the Economics of Aggregation, I argued that market structure systematically underprices aggregators like Alphabet. Since then, the multiple has drifted up - from 20x P/OCF in October 2024 to about 22.5x P/OCF today.

I value Alphabet on operating cash flow (OCF), not free cash flow or net income, because the company is in a CapEx super-cycle (front-loading AI and cloud infrastructure across servers, TPUs/GPUs, and data centers), which suppresses near-term FCF and net income is distorted by volatile Other Income and large non-cash D&A. OCF best reflects the business’s true cash generation in this phase.

Trailing twelve months (TTM) OCF per share rose from $8.46 (Q3 2024) to $12.41 (latest), +46.62%, while revenue per share climbed to $31.59 (+15.4% YoY). Over the same span, the stock advanced to $281, up ~65% since October 2024. The move from 20x to 22.5x P/OCF wasn’t multiple expansion alone; it came with meaningful OCF/share acceleration as Google widened its AI lead and the CapEx cycle began to monetize.

Now, what does that cash engine imply five years out? Using TTM OCF of $151.4 billion as of September 30, 2025 as a base, assuming a conservative growth rate of 12% annually for five years gets to $266.86 billion. Let share count shrink 2% a year (the share count net reduction was -2.31% yearly for the past 5 years) to 11 billion shares on buybacks. Hold the exit P/OCF at 22x and you get a future stock price of $533 ($266.86 billion / 11 billion shares x 22). Add dividends ($0.21 quarterly payout2 is $0.84/year or $4.20 over five years) and total future value is $536/share. Discounted back five years at 10%, that yields a today fair value of $333/share, a 18% discount to today’s price of $281 and a 5-year compounded annual growth rate (CAGR) of 13.8%, above my 10% hurdle rate.

22.5x P/OCF looks full versus the 18x 10-year average, but the trajectory matters. Services is sturdy; excluding EC fine, margins are rising, and AI-native upgrades in Search/YouTube/Workspace should lift revenue and unit economics as ad formats mature. Cloud at 23.7% margin is early; as on-prem shifts to cloud and mix tilts to enterprise AI infra/solutions, each margin point disproportionately boosts consolidated profit. If AWS’s 30-35% is a guide, 30%+ for GCP is a reasonable long-run target. CapEx is dampening FCF now but is rational if ROIC stays high; backlog, deal mix, and surging AI token volumes support that.

Risks remain: execution must match CapEx intensity; supply constraints could bite; and regulatory overhangs, from fines to remedies, can pressure margins and strategic optionality. Competitive pressure, particularly from Azure’s deep enterprise embedding, isn’t going away.

Net-net, despite the stock’s run, the market still isn’t fully pricing the compounding from Alphabet’s full-stack AI position, cloud margin runway, and resilient Services economics. On my scorecard, Alphabet remains undervalued and it continues to be a core holding in my portfolio.

Final Thoughts

Alphabet just delivered a quarter that reconciles the narrative with the numbers. Search, YouTube, and Subscriptions are compounding; Cloud is scaling with improving unit economics; and AI is no longer a press release, it’s a P&L driver. The company is spending aggressively on infrastructure, but the return signals are already visible in Cloud backlog, rising Cloud margins, and the adoption of AI-native surfaces across the core products.

I anchor on operating cash flow. Despite the multiple drifting from 20x to 22.5x P/OCF, the business has outgrown the multiple: OCF/share rose 47% over the past year, and a five-year build from today’s OCF, conservative share count shrink, and a steady exit multiple still points above today’s price even after a fresh rally. Add the optionality (Waymo, AI agents, and other bets), and the thesis remains simple: the market continues to underprice a full-stack AI platform with durable distribution and accelerating monetization.

I’m offering a fundamental perspective here: I don’t own CoreWeave (or any of the so called neoclouds), and as a value investor I care about understanding what I’m buying and the dynamics of the industries I invest in, not chasing momentum.

After studying the S-1 filing1, the financing structures, and the synthetic demand loops, my takeaway is sober: CoreWeave is highly speculative, and its model (heavy leverage, concentrated customers, fast-aging assets) raises red flags that deserve scrutiny from anyone considering the stock.

The Bull Case: Explosive Growth, Hyperscaler Demand, and Its Hidden Risk

Hyperscalers like Microsoft, Google, and Amazon face acute compute shortages, opening the door for data center operators such as CoreWeave to rent GPU capacity and act as “shadow clouds” for AI workloads. This outsourcing model has produced staggering top-line growth across neoclouds, but much of that demand is really a transfer of balance-sheet risk: hyperscalers expand capacity without tying up their own capital, while debt-financed neoclouds assume the depreciation, interest costs, and refinancing pressure. Great for Big Tech’s CapEx optics; capital-intensive and fragile for the neoclouds.

Why It’s Risky, in One Glance

• Negative cash generation: H1 2025 operating cash flow was negative; debt service was funded largely by new borrowings and IPO proceeds, not operations.

• Refinancing wall: $986 million due in 2025 and $4.2 billion in 2026 must be refinanced or repaid, likely at high rates.

• Shrinking collateral: Loans are secured by GPUs that depreciate quickly as new NVIDIA chips launch; principal stays fixed while collateral value falls.

• Covenant exposure: Must meet leverage and contracted-revenue tests; a miss can trigger paydowns or tighter terms.

• Customer concentration: Microsoft was 62% of 2024 revenue (Q2 2025’s Customer A, 71%). Any pullback risks breaches right as collateral ages.

• Ongoing external funding: Management says growth needs more debt/equity, meaning higher interest costs or shareholders’ dilution.

From Crypto to Cloud: CoreWeave’s Reinvention

Founded in 2017 as a crypto miner under the name Atlantic Crypto, CoreWeave pivoted to GPU rentals and then to cloud compute for AI. After generative AI took off in late 2022, it landed major contracts with Microsoft and OpenAI.

Today, it reports 250,000+ NVIDIA GPUs across 33 data centers with 470MW of active power - among the largest GPU fleets outside AWS, Azure, and GCP.

They’ve been dubbed “the multibillion-dollar backbone of the AI boom” by Wired, but backbones don’t usually run on floating-rate debt, negative operating cash flow, and a single customer supplying most of the revenue.

Financials: Rapid Growth, Rising Losses

When CoreWeave filed its S-1 filing earlier this year (targeting $35 billion IPO valuation), it warned:

The filing also disclosed material weaknesses in internal controls and extreme customer concentration.

Key numbers:

• 2024: $1.9 billion revenue; GAAP operating income $324 (17% margin) and GAAP net loss $863.4 million.

• Q2 2025: $1.2 billion revenue; GAAP operating income $19.2 million (2% margin) and GAAP net loss $290.5 million.

• Funding: $14.5+ billion total, mostly debt.

• CapEx: $8.7 billion in 2024 (YTD $3.9 billion, $2.9 billion in Q2 2025 alone).

• Interest expense: $349.2 million in 2024 (YTD $513.5 million) - it may exceed $2 billion/year.

The company reported $1.2 billion in revenue in Q2 2025, with a $30 billion backlog. Yet, despite a 62% adjusted EBITDA margin, GAAP operating income was only $19.2 million (2% margin) and GAAP net loss was $290.5 million in the quarter.

The reconciliation explains why: The two items adjusted EBITDA ignores are huge - $559.5 million in depreciation & amortization (46% of revenue) and $267 million in interest (22% of revenue) and they’re not accounting quirks; you can’t ignore GPU decay or debt service and call it “profit”. Add $2.9 billion of Q2 2025 CapEx, and it’s obvious why EBITDA is the wrong yardstick for a CapEx-heavy, fast-obsoleting GPU fleet: with GPUs turning over every 3-4 years, depreciation is a recurring cost of staying competitive, not a non-cash nuisance. Treating adjusted EBITDA as earnings here is like valuing a steel mill while pretending the furnaces never wear out.

Concentration risk is rising, not falling. 62% of 2024 revenue came from Microsoft; another big slice appears to be NVIDIA renting back its own GPUs. In Q2 2025, Customer A (likely Microsoft) was 71% of revenue, with the next three customers (two of which are likely OpenAI and NVIDIA) each <10%.

Even with scale, profitability doesn’t show up. This isn’t the profile of a thriving, diversified platform. It’s a company built on sand, propped up by a couple hyperscaler clients and a debt-fueled, synthetic demand cycle.

Debt & Leverage: CoreWeave’s Debt-Fueled Growth Machine

CoreWeave runs on borrowed money. By June 30, 2025, total debt had ballooned to $11.17 billion, up from $8 billion just six months earlier - excluding another $3.5 billion in lease liabilities. Most of this sits in high-interest, secured term-loan facilities tied to depreciating GPUs, so the very hardware generating revenue also underwrites rising leverage.

The bulk of CoreWeave’s financing comes from its Delayed-Draw Term Loan (DDTL) facilities - effectively large, asset-backed loans that can be tapped in stages to fund GPU purchases. DDTL 1.0, arranged in 2023 for up to $2.3 billion, carries a floating rate of roughly 15%, amortizes quarterly, and comes due with a balloon payment in March 2028. It is guaranteed by CoreWeave and secured by its subsidiaries’ assets, with strict prepayment penalties.

In May 2024, the company added DDTL 2.0, a much larger $7.6 billion facility with a variable interest rate averaging around 11%, determined by the credit quality of customer contracts backing each draw. By June 30, 2025, CoreWeave had drawn $4.95 billion and extended the draw window to September 2025. Repayments begin in January 2026, with each tranche maturing five years after funding.

In July 2025, CoreWeave introduced DDTL 3.0, a new $2.6 billion facility to finance infrastructure for a “strategic customer”. It’s priced at a floating benchmark plus 4%, includes a 0.5% annual fee on undrawn funds, begins amortizing in 2026, and matures in August 2030.

These DDTL loans are the engine of CoreWeave’s growth, but also its Achilles’ heel. They’re collateralized by GPUs whose market rental rates have already fallen 50-70%, shrinking collateral value just as repayment schedules begin referencing depreciated hardware. The company itself proudly highlighted this structure in its S-1 filing:

The structure may sound sophisticated (secured, self-amortizing loans tied to specific customer contracts) but it’s fragile. In theory, each loan’s principal and interest are covered by customer cash flows during the contract term, supposedly producing “positive free cash flow”. In reality, the company is borrowing against assets that must be replaced every few years, so much of that “cash generation” is already committed to the next CapEx cycle. It’s less a flywheel than a treadmill - constant motion without real progress in cash generation.

Through the first half of 2025, CoreWeave paid $2.088 billion toward debt service - $1.575 billion in principal and $513.5 million in interest, while posting a $190 million operating cash flow loss.

With total debt up 40% in six months, climbing to $11.17 billion, and leverage above 70%2 of total capital, the balance sheet is running on financial momentum, not operating cash.

At an average borrowing cost near 11%, CoreWeave’s annualized interest bill now approaches $1.2 billion, or roughly $100 million/month, before factoring in new obligations such as the $2.6 billion DDTL 3.0 and $2 billion in 10% senior notes issued mid-2025. Including these, total monthly interest could soon exceed $120 million, depending on rate resets and drawdowns. Principal repayments add another $250-300 million/month, implying total monthly debt service of $350-420 million, far beyond what the business generates from operations.

As of mid-2025, cash and equivalents (including restricted) stood at $2.05 billion, with roughly $4 billion in remaining borrowing capacity, much of it fenced off by covenants, escrow requirements, and usage restrictions.

At its core, this is a high-debt, high-CapEx model built around a few hyperscaler clients and rapidly aging hardware. In perfect conditions - strong AI demand, easy credit, and cooperative lenders - it can function. But if demand softens, credit tightens, or a major customer hesitates, the same leverage that fuels CoreWeave’s expansion could quickly become its choke point. Until the company produces durable, positive operating cash flow, its debt looks less like growth capital, and more like a ticking constraint.

A Commodity Business Dressed as a Platform

CoreWeave rents GPUs by the hour, a commodity service with little differentiation. Unlike AWS, Azure, or GCP, which layer profitable software and services’ ecosystems atop their infrastructure, CoreWeave lacks a sticky developer platform, proprietary tools, or meaningful brand moat. Its only advantage is proximity to NVIDIA and preferential chip access, an edge that can erode quickly as others scale up. In a race to install the most GPUs and rent them out, price becomes the lever, margins compress, and the business drifts toward the bottom-line economics of a commodity.

The company presents this hourly rental and contract model as a proven, repeatable engine of growth, yet in reality, the company has only recently completed its first full 54-month lifecycle since pivoting from crypto mining to AI data center (2019-2020). The slides depict a mature, steady cash-flow machine - 3-month install, then monthly payments in months 6-30 and 30-54, with GPUs rolling into an on-demand pool afterward. With its AI-specific scale-up only starting in 2023, CoreWeave has barely two years of track record under this model.

At the Deutsche Bank 2025 Tech Conference on August 27, 2025, Nitin Agrawal, the company’s CFO reinforced this framing, saying CoreWeave still evaluates projects on a 2.5-year payback (on an adjusted EBITDA basis), consistent with the S-1 filing: secured, self-amortizing DDTL debt is tied to specific customer contracts and structured over five years, with the goal that interest and principal are covered during the contract term while throwing off positive free cash flow to the company.

That may be true on paper, but the economics still deserve scrutiny. A 2.5-year EBITDA payback is not a profit payback; it ignores depreciation/obsolescence and the fact that GPUs typically need replacing every 3-4 years. By the time a primary term ends, CoreWeave faces another large CapEx cycle just to stand still. And because the loans are collateralized by depreciating GPUs, the collateral base shrinks even as replacement spending looms. In short, the inflows during months 6-54 look less like durable profits and more like temporary cash returns in a perpetual replacement loop, one that depends heavily on a few large contracts to keep refinancing and rollout on track.

These structural weaknesses feed directly into NVIDIA’s ecosystem.

The NVIDIA Loop: How the GPU Giant Became the Engine of AI Infrastructure

NVIDIA has evolved beyond supplying GPUs to become a key architect and enabler of AI infrastructure. After acquiring Mellanox Technologies in 2020, NVIDIA expanded its networking and data center stack and has seen its Data Center segment become the biggest revenue and profit driver. In parallel, NVIDIA has invested in, or partnered with, AI data centers like CoreWeave, Nebius Group (), and Lambda Labs structuring deals where the neoclouds borrow against long-term contracts with hyperscalers to purchase NVIDIA GPUs. In some cases, NVIDIA steps in as a safety valve, for example, an initial order to lease unsold compute capacity (e.g. a $6.3 billion agreement in September 2025 with CoreWeave). The loop amplifies GPU sales and installed base, but it also financializes demand: growth depends on credit markets, contract durability, and rapid hardware refresh cycles. CoreWeave, with 470MW active vs. 2.2GW contracted, still must finance vast expansion, mostly with debt collateralized by depreciating GPUs. It’s a fragile flywheel powered by leverage that can keep spinning only while credit and chip cycles cooperate.

Final Thoughts: Revenue Is Not Shareholder Value

The CoreWeave story captures both the promise and peril of the AI infrastructure boom. Behind the growth headlines lies a business model built on borrowed money, depreciating hardware, and a handful of customers. For now, debt and hype mask the fragility, but the economics eventually surface - they always do.

None of this means CoreWeave can’t succeed or that the AI build-out will stop. It means that level-headed investors should separate AI as a secular trend from the companies financing it. Leverage, concentration, and obsolescence make for a difficult mix even in expansionary cycles, and a dangerous one when liquidity tightens.

To borrow Jeff Bezos’s framing, AI’s trajectory resembles an industrial bubble: the technology’s promise is real, but the financing mechanisms around it are fragile. Like the fiber-optic boom that bankrupted its builders yet connected the modern internet, the AI buildout will likely leave behind lasting infrastructure - even if some current players don’t survive. That’s the paradox of industrial revolutions: society gains enduring assets, but many investors don’t.

As a value investor, I’m drawn to businesses that compound capital from within, not ones that must constantly borrow it. In CoreWeave’s case, the question isn’t whether demand for GPUs will grow, but whether the company can ever capture durable economics from it. Until it proves that with consistent operating cash flow and a self-funding model, the risks remain greater than the rewards.

For more than three decades, Brookfield has compounded shareholder value at 19% annually - that would have turned $1,000 into $270,000 in 30 years.

As Benjamin Graham put it, “In the short run, the market is a voting machine; in the long run, it is a weighing machine”. The 3% pop around Investor Day was the voting; three decades of disciplined compounding is the weighing - real cash flows, durable assets, and steadily expanding base of fee-bearing and perpetual capital being valued over time. As CEO Bruce Flatt put it:

That consistency comes from a long-term investment philosophy that blends value investing, operational excellence, and methodical capital allocation. It echoes Jeff Bezos’s idea point about ignoring short-term market applause to build a “heavy” company. I see it in Brookfield’s growth engines: fee-related earnings and carry from Asset Management, insurance float compounding within Wealth Solutions, and the continued expansion of real asset platforms, such as infrastructure, power, and real estate, in both operations and financial weight. Over time, that accumulated mass pulls intrinsic value, and eventually market price, higher.

Management framed the company’s next chapter as a transformative phase of growth, powered by three global trends:

• AI-driven infrastructure investment.

• Aging demographics fueling new wealth and retirement products.

• A global real estate recovery.

Brookfield is targeting 25% compound annual growth (CAGR) in Distributable Earnings (DE) per share and 16% annual plan value1 growth from $102/share to $210/share (pre three-for-two stock split) by year-end 2030; goals the team reiterated throughout the day.

Finding Trends, Packaging Products, Deploying at Scale

Brookfield’s innovation playbook is disarmingly simple: spot global secular trends, productize them for clients, then deploy capital at scale. As Flatt put it, they “identify the trends on a big, global basis […] package them into products for our clients, which then enables us to attract scale capital within each sector”.

Crucially, Brookfield doesn’t try to be the earliest; it aims to be fast once it understands an industry. That readiness-to-scale is a competitive edge: “We’re not early, but once we understand it, we move fast, and we deploy at scale”. It’s an approach that has repeatedly expanded the company’s opportunity set, “50% of the assets that we operate with today didn’t exist as an asset class for investible purposes 15 years ago”.

Even as Brookfield widens into adjacent strategies to meet differing return profiles, the investment approach stays consistent: “operate with a value investing thesis taking moderate risk, aligning ourselves with our constituents, using our operating skills to squeeze extra returns out of it […] a methodical and proven capital allocation process […] and, probably most importantly, the culture we have in the organization is extremely important”.

AI Factories: Power, Property, and Patient Capital

Brookfield sees AI infrastructure as a generational opportunity — the new railroads, highways, and water systems of the next decade. The company has identified seven large-scale AI factory sites, representing 6GW of compute and ~$200 billion of investment across North America, Europe, and the UK.

Flatt’s remarks put a point on the scale and runway:

Given hyperscalers’ power constraints and Brookfield’s integrated capabilities (power, transmission, real estate, and capital), the business is positioned to be a critical enabler of AI infrastructure for years.

Bringing Alternatives to Individuals

Brookfield highlighted a rising pool of capital beyond institutions and its own balance sheet: individual investors (401(k), annuities, private wealth). Flatt described both the size and the shift:

Brookfield plans to capitalize on this shift through new products designed to bring the benefits of private investing to individuals, generating a long runway that compounds fee-bearing capital and deepens client relationships. The business’ decades-long track record managing private assets positions it as a natural leader in this emerging space. This segment is part of Brookfield Wealth Solutions, discussed below in this article.

Subscribe now

Real Estate: Playing Offense Into the Recovery

Real estate has been one of Brookfield’s defining competencies. While many peers have pulled back amid tighter credit and higher rates, Brookfield continued to deploy capital, roughly $60 billion over the past five years, across logistics, multifamily housing, hospitality, and life sciences properties.

The team emphasized that fundamentals are firming and transaction activity is returning. Kevin McCrain, who leads Real Estate, was direct:

Within the portfolio, Brookfield now segments balance sheet’s real estate into Super Core, Core Plus, and Value Add. The Super Core group (34 premier assets) runs at 95% occupancy with $19 billion of equity and the Core Plus set (57 assets) runs at 94% occupancy and carries $8 billion of equity.

On the ground, leasing and rent trends are tightening as new supply remains constrained and return-to-office policies normalize:

The company believes this disciplined activity leaves it in a position of “unmatched strength” as the next cycle begins. Real estate, historically cyclical, has often been a source of outsized returns for patient allocators willing to invest when sentiment is poor, and that approach fits perfectly within Brookfield’s contrarian, value-oriented DNA.

BWS: Building an Investment-Led Insurance Flywheel

One of the standout presentations came from Sachin Shah, CEO of Brookfield Wealth Solutions (BWS), who outlined why Brookfield built this business five years ago, and why it’s now central to the company’s future.

The premise starts with demographics: an aging population across the Western world, longer lifespans, shrinking birth rates, and the erosion of defined-benefit pension plans. A century ago, only one in twenty Americans were over 65; today, it’s one in six, and by 2050, it’s projected to be one in four.

This shift, combined with a $7 trillion U.S. retirement income gap, is driving enormous demand for reliable income products.

Brookfield saw this as both a societal challenge and a generational investment opportunity. The idea was to marry its decades-long expertise in managing long-duration, cash-generating real assets with the stable, predictable capital base that comes from insurance float (the policyholder capital it invests on behalf of clients). As Shah put it, Brookfield could “compound capital for many decades to come” by aligning its investment strengths with the liability needs of retirees seeking steady income.

Initially focused on reinsurance, the strategy evolved toward a fully integrated insurance platform, owning its own distribution, writing its own policies, and directly matching long-term liabilities with high-quality, long-duration investments in infrastructure, power, real estate, and credit. This model not only fits Brookfield’s DNA but also complements its asset management arm (BAM), which earns fees for managing these assets on behalf of BWS, creating a symbiotic investment-led insurance flywheel.

In essence, Brookfield is using its investment expertise to turn insurance float into a high-return compounding engine. It’s a modern reinvention of the Berkshire Hathaway playbook, one that matches Brookfield’s global scale and long-term mindset. The philosophical shift from traditional to investment-led insurance is deliberate. As CEO Flatt explained:

Most insurers aim to profit from underwriting (pricing policies so that premiums exceed claims) while investing the resulting float conservatively to avoid losses. Brookfield flips that logic. It keeps underwriting low-risk and roughly breakeven because the real goal is to invest the float in long-duration, cash-generative assets it already knows well. BWS only grows when there are attractive investment opportunities for that float, maintains ample liquidity for safety and regulatory compliance, matches long-dated liabilities like annuities with long-term assets, operates with lower leverage, and targets 15%+ returns on the equity (ROE) supporting the business.

A simplified example illustrates the model. Suppose BWS writes $1 billion of annuities. To back those promises, it holds $1 billion in assets - say, $500 million in liquid high-quality bonds and $500 million in BAM-managed real assets. If the bonds earn 5% and the real asset sleeve earns 12%, the blended portfolio return is 8.5%. Credit customers receive 5% on their annuities; insurance operations cost 0.5%, leaving a 3.0% spread, about $30 million a year. Regulators require roughly $150 million of Brookfield’s own equity to support that $1 billion block, so $30 million of annual profit on $150 million of equity implies 20% ROE. A modest spread on a large float base can translate into high returns on the smaller slice of Brookfield’s own capital at risk.

BWS is exactly the synergistic counterpart Brookfield envisioned when it spun out BAM. In Flatt’s framing, the goal isn’t scale for its own sake, but durable float that can be invested through BAM. BWS sources patient, low-risk liabilities; BAM converts them into long-duration, cash-generative assets. That closed loop is the flywheel.

BWS has already scaled from zero to $1.7 billion of DEs in five years, and now manages $135 billion in insurance assets (total balance sheet capital held across its insurance and reinsurance subsidiaries). Management expects it to contribute over one-third of Brookfield’s per-share profit growth in the coming five years and potentially generate more than $5 billion in annual DEs by 2030 - more than the entire company produces today.

Crucially, BWS has been built with regulatory rigor and balance sheet prudence. It operates with excess capital, low leverage, transparent disclosure, and strong regulator relationships, what Shah calls “a moat around the business”. With platforms now spanning the U.S., Canada, the U.K., and a developing foothold in Japan, Brookfield’s insurance footprint has become truly global. Following its recent U.K. acquisition, BWS now oversees $180 billion of insurance float and expects $20-25 billion of net inflows each year. BWS is well positioned to benefit from the global demand for retirement income solutions.

BAM: The Compounding Core (Fees + Carry)

Brookfield Asset Management (BAM) remains the engine at the center of the ecosystem. The business unit continues to grow its fee-bearing capital base and expects 20% annual earnings growth through 2030. By that year, DEs are projected to reach $7.7 billion, driven by both recurring management fees and realized carried interest. Goodman underscored that carried interest is entering a new phase:

Despite this, the market seems to underestimate the value of carried interest, a recurring theme across alternative asset managers. Brookfield’s historical ability to exceed its own five-year projections, however, lends credibility to its forward outlook.

BAM targets 20% annual earnings growth through the plan period, driven by fee-related earnings, carry realization, and scaled multi-strategy fundraising.

5-Year Track Record, Strong Numbers

Goodman’s summary of the last five years was straightforward: “we have delivered earnings growth of 22% […] and planned value growth of 16% on an annualized basis”, exceeding the 2020 Investor Day targets the company set for itself. He added that the “growth profile of the outlook […] is very strong,” underpinned by a conservative balance sheet and ample liquidity.

Across the business, capital formation and recycling accelerated: Brookfield “raised over $95 billion of capital […] deployed $135 billion […] monetized over $75 billion of assets” over the last 12 months, a step-change from prior years. Plan value rose to $102 per share (pre-stock split) with a 22% increase in the past year alone; proof, in Goodman’s words, of “$28 billion of value added in just 12 months”. Brookfield also returned “over $1.5 billion” to shareholders (dividends plus opportunistic buybacks), emphasizing that repurchases at discounts “added value to each remaining share”.

Asset Management (BAM)’s fee-related earnings grew 18%, supported by broad-based fundraising where “75% of [capital came] from complementary strategies” reflecting the platform’s diversification.

Wealth Solutions (BWS)’s DEs reached $1.7 billion while “maintain[ing] returns on equity in line with our long-term targets of 15%+”.

In the operating businesses unit, renewables and infrastructure grew Funds From Operations (FFO, akin to operating cash flow) by 13%; Real Estate “monetized $5 billion” from the balance sheet; Essential services & cash-flowing private equity businesses lifted EBITDA 15%.

Targets beaten, base strengthened - Goodman reminded investors that the 2020 plan to reach $4.8 billion DE before realizations and $100 plan value per share by 2025 was surpassed:

Over the same 5-year period, the stock price followed suit with a 17% total compound annual return on plan value and with 19% over 30+ years.

Share

Price-Value Gap and the 5-Year Path

Goodman reiterated that Brookfield’s valuation approach “remains unchanged”: sector-appropriate methods summed the parts to $161 billion of plan value ($102/share pre-stock split) net of debt and preferreds. Despite the past year’s rally, he called today’s level “a very, very attractive entry point”: on Brookfield’s projected 5-year DEs, the stock is trading today at 10x forward DE.

As of September 9, 2025, Brookfield’s share price was $66.2, a 35% discount to management’s plan value of $102/share. I came to a slightly lower valuation on February 18, 2025 in a previous article based on Brookfield’s 2024 financial results and 5-year outlook at the time, as the financial results did not factor in the growth that the company has experienced in H1 2025.

Brookfield sees the stock price to intrinsic value gap closing through two engines: core growth (Asset Management, Wealth Solutions, operating businesses) and capital allocation (recycling, buybacks, strategic deployment). Management’s base case targets 25% annual DE/share growth over the next five years - 20% from the business units and 5% from capital allocation.

On that path, DE/share is modeled to reach $6.90 before realizations/capital allocation and $10.40 in total by 2030, while plan value compounds 16% from $102/share ($68/share post-split) to $210/share ($140/share post-split).

Why this outlook is credible: carried interest is at an “inflection point” (management expects $25 billion of net realized carry over the next decade; $6 billion within three years), free cash flow is projected at $53 billion over five years with $25 billion of optionality after dividends and organic reinvestment, and the balance sheet remains conservatively capitalized to support recycling through cycles. Add the strategic backdrop: rising allocations to alternatives from institutions and individuals, the scaling integration of BWS, AI infrastructure investment, and a real-estate recovery, and the investment case becomes compelling.

Brookfield has beaten its prior 5-year targets, and management aims to repeat that cadence with a disciplined allocation flywheel and increasingly visible carry. If Brookfield delivers $10.40 in DE/share (pre-stock split) by year-end 2030, the forward DE multiple on pre-stock split’s $102/price is 9.8x, a compelling setup reinforced by the current price-value gap. From today’s post-split price of $44/share to a projected post-split plan value of $140/share, the implied 5-year return is a 26% CAGR.

Final Thoughts

Brookfield combines the qualities I value most: owner-oriented leadership, a disciplined capital allocation framework, strategic adaptability, and a long runway of secular growth. It isn’t about short-term excitement; it’s a long-term compounding machine that keeps evolving, from traditional infrastructure to renewables and AI data centers, an investment-led insurance platform, and broader wealth solutions. Investor Day reaffirmed the vision and the execution discipline behind that evolution. The targets are ambitious, but so is the track record; few global platforms are as diversified, aligned, and forward-looking.

What stands out is Brookfield’s ability to add new, scalable profit pools without losing rigor, spinning out listed entities (2008), launching Wealth Solutions (2020), spinning off BAM (2023), and now building AI factories. As Bruce Flatt put it:

That mindset, paired with management’s deep alignment and multi-decade tailwinds in AI, wealth, and real estate, is why I continue to view Brookfield as a long-term compounder worthy of a top position in my portfolio.

For a deeper dive into each business unit, I recommend watching the info-packed Investor Day recordings on Brookfield’s YouTube channel.

Subscribe now

Entrepreneurship and Value Investing Principles

Bezos’ framework merges founder mentality with investor discipline. The mindset starts with optimism used productively, not blindly. Bezos frames optimism not as naiveté but as fuel for for exploration under uncertainty: the “delusion” isn’t about ignoring risk; it’s about having enough forward energy to do hard, compounding work before the data is fully conclusive.

From mindset to method, the first principle is invariants, the few customer needs that barely move across decades. Bezos’ north star hasn’t moved in 30 years: speed, price, reliability. Those are as investable today as they were in 1997. Enduring strategy anchors to those needs and flexes the “how” as technologies and competition change.

This is classic value investing logic applied operationally: underwrite to what won’t change; iterate everything else.

From there, the discipline is to build weight, ignore votes. Bezos rewound to the dot-com crash to show the gap between price and progress. In 2000, Amazon’s stock fell from $113 to $6 in short order; anxious shareholders, nervous employees, parents calling to ask if the company would survive. That was the voting machine at work. Yet month after month, as the stock price collapsed, core metrics improved: more customers, rising gross profit, losses shrinking as a percentage of sales, stronger repeat purchasing. In his words, “every single business metric […] kept getting better”. Stock prices detached from fundamentals during the bubbles. Hence Benjamin Graham’s touchstone Bezos cited:

For founders and investors, the practical translation is manage to fundamentals, not to mood, treat the stock price as an output you don’t control, echoing Amazon’s 2009 shareholder letter, where Bezos wrote:

Bezos then gave the operating mandate that ties the idea together:

“Heavy” here is not mystical. It’s the accumulation of customer trust and behavior (acquisition, retention, repeat use), unit economics (gross profit, contribution margins), and operating discipline (cost as a share of sales, cycle times, defect rates). When those inputs get denser over years, the weighing machine does its work, on its timetable, not ours. For a deeper link to Graham, the Mr. Market allegory in The Intelligent Investor book explains why the voting machine can swing wildly while fundamentals are quietly weighed - you can read my earlier article on this subject.

Invention then needs pacing: wander to discover; regulate the release rate to deliver. Great invention looks “inefficient” on the way in and “inevitable” on the way out. Exploration is necessary, but it must be paced so the organization can absorb it.

Bezos recalls the time a former Amazon senior executive, Jeff Wilke, told him:

The lesson is to backlog ideas, sharpen the few that matter, and release them at a cadence the business can execute; operational cadence is part of the edge. That cadence depends on who you keep in the company. Mavericks are an asset class, defend them. Most large companies develop antibodies to the very people who can spark the next cycle of innovation and growth.

Adaptation closes the loop. Advantage accrues to the fastest updater, in products and in theses. If a thesis can’t evolve with evidence, it isn’t a thesis, it’s a slogan.

Finally, the compounding substrate is trust. Culture becomes cash flow over decades; choosing the relationship over the transaction creates the permission to reinvest.

Operationalized as reliability, fairness, responsiveness; kindness reduces churn and raises lifetime value - the quiet mathematics behind durable moats.

AI, Bubbles, and What Actually Matters

Bezos separates market cycles from capability shifts. The core claim is unambiguous:

Bezos calls AI a “horizontal enabling layer” whose biggest impact won’t be a handful of AI-first startups, but the quiet upgrade of quality and productivity across the economy:

The timeline is uncertain, diffusion will arrive at different speeds by sector, yet the direction is “very real”.

Hype cycles, he notes, have familiar mechanics: when enthusiasm peaks, capital funds almost everything:

Valuations can detach from fundamentals for long stretches. That doesn’t invalidate the underlying technology; the market just misprices the journey. To explain why society often benefits even when investors don’t, Bezos distinguishes industrial from financial bubbles. The former can seed enduring assets:

In the 1990s biotech boom, “as a group [investors] lost money but we did get a couple of life-saving drugs.” The 2000s dot-com era offers the same lesson:

Hence his summary:

What, then, actually matters for founders and investors during AI’s boom? First, decouple sentiment from substance. Treat stock price as an output; monitor the inputs (customer adoption, quality gains, unit-level productivity). Second, understand where the value will accrue: not only to new AI vendors, but to operators who integrate AI to raise reliability, speed, and cost efficiency. As he puts it, the diffusion is economy-wide:

Finally, I keep the long arc in view. Technological progress compounds into broader prosperity:

Amid the noise and volatility, the signal is this: AI’s benefits are durable, the infrastructure will outlast the market cycle, and the weighing machine will eventually reward the businesses that turn real capability into real productivity. In today’s context, that vector points to the hyperscalers, especially full-stack providers (silicon, cloud fabric, model layer) like Amazon and Alphabet; I lay out that thesis in an earlier article.

Final Thoughts

Bezos has a gift for focusing on what matters. He’s an exceptional founder and business operator, and a charismatic speaker I continue to learn from. He keeps returning to first principles: customer invariants, disciplined execution, and long-term alignment. I highly recommend watching the full recording of his conversation with John Elkann: beyond the themes I covered here, he recounts his youth and upbringing, Elkann speaks vividly about his grandfather’s influence, and Bezos extends his vision to space, what Blue Origin is building, why it matters for humanity, and the hard technical problems they’re solving.