Invert the Market's AI CapEx Panic
Why "supply constrained" is the tell and what would prove returns on incremental AI capacity
“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:
“If you own a stock in a company and they announce that maintenance CapEx is going to be twice what we thought it was going to be […] the stock should go down a lot.”
But when the spend is driven by demand, the reaction should invert:
“If a company announces that business demand is so great that in order to meet the demand, they need to build more factories […] you should applaud when they’re doubling the amount of money that they’re spending.”
Their punchline is exactly the one the market often misses:
“[…] we think this is squarely in the growth CapEx category as opposed to the maintenance CapEx category.”
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?
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:
“If you are already an industry leader […] if you have the distribution, if you have the data, if you already have a large degree of competitive moats […] these businesses have a unique moment right now to further widen their moats”
That lines up with the thesis underneath both of my earlier articles:
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.
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.
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.
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:
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.
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.