The Buildout Bet

1. The week the cash machine blinked

On July 22, 2026, Alphabet did something it had never done in the quarter-century it has been the most dependable cash generator in corporate history: it spent more building for the future than its entire business threw off in cash. Second-quarter capital expenditure came in at $44.9 billion, roughly double the $22.4 billion of the same quarter a year earlier, and free cash flow turned negative for the first time on record [1, 2]. Management did not flinch. It raised full-year 2026 capex guidance to $195–205 billion, up from a prior $180–190 billion range that had itself already been lifted once, and funded the acceleration in part with $49.6 billion of fresh equity earmarked for AI infrastructure [1]. The market's verdict was immediate and sour; the stock fell even as revenue beat.

The tell was not the number. It was the confidence. Google Cloud revenue grew 82% year over year to $24.8 billion at a 35.6% operating margin, monetization that is real, accelerating, and margin-accretive rather than bought at a loss [1, 2]. Alphabet's chief financial officer, Anat Ashkenazi, framed the spending in the language of a firm chasing a line it cannot reach: "We have increased our capacity quite significantly over the past three years. The demand still outpaces that investment" [1]. A company that could coast is instead borrowing and issuing stock to pour concrete and buy silicon, and it is doing so because customers are, by its own account, waiting in line.

That is the shape of the question this column exists to answer. When the most disciplined operator in the industry runs its own cash reserves negative to keep building, one of two things is true. Either the demand is genuine and the scarcity is binding, in which case the spending is the rational response of a firm that would lose the decade by underbuilding, or the demand is partly a mirage sustained by circular capital, and the industry is marching, in lockstep and on borrowed money, toward the largest write-down in the history of technology.

My answer, which the rest of this essay will earn rather than assert, is that both camps are half right, and that the half each gets wrong is the same half. The compute is a rational asset to own. The way it is being financed and depreciated is where the bubble lives. If the buildout ends badly, it will not end because nobody wanted the chips. It will end because the paper stacked on top of the chips (the six-year depreciation schedules, the vendor loops, the debt moved off the balance sheet) promised a smoother and sooner return than the physics of the asset can deliver.

2. The size of the bet

Start with the magnitude, because the magnitude is the reason the debate is not academic. The five largest hyperscale spenders — Amazon, Alphabet, Microsoft, Meta, and Oracle — are now widely forecast to exceed $600 billion in combined capital expenditure in 2026, a 36% increase over 2025, with roughly $450 billion of that, about 75%, tied directly to AI infrastructure rather than to traditional cloud; the underlying cloud-infrastructure data comes from Omdia (Tier B, analyst forecast, December 2025) [3]. The individual guidances stack toward the same total: Amazon near $200 billion, Meta $115–135 billion, Microsoft above $120 billion, and Oracle around $50 billion for the year, alongside Alphabet's $195–205 billion (raised in July, as §1 describes; the earlier compilations still carried a pre-raise $175–185 billion) (Tier C, compiled analyst estimates) [4].

Two features of this spending matter more than the headline. The first is where the money goes. Chips and the servers that carry them dominate the bill; by Patel and Dean's reckoning, chips alone are "an overwhelming 60+% of the cost of a data center" [23]. The physical composition of an AI data center is unusual: the compute is the overwhelming majority of the cost and the fastest-depreciating part of the whole, a point the buildout's defenders and its critics agree on before they agree on nothing else.

There is a threshold this spending has now crossed. Alphabet's own quarter shows it: a company that could comfortably self-fund instead ran free cash flow negative and raised $49.6 billion of equity earmarked for AI infrastructure [1, 2]. When the most cash-rich operator in the group must reach outside its own cash flow to sustain the pace, the marginal dollar of buildout becomes an externally funded dollar rather than a self-funded one. That shift is what turns a spending debate into a solvency debate, and it is why the financing structure, not the spending level, is where this essay eventually lands.

The second feature is the depreciation tail that trails the spending. Every dollar of this capital lands on the balance sheet as an asset that must then be written down over its useful life, and on a base being expanded this fast the annual depreciation charge becomes one of the heaviest recurring costs the buildout carries. That charge is the hinge of the entire argument. Depreciation is not an accounting footnote here; it is the mechanism by which a capital-expenditure boom either converts into durable earnings or reveals itself as a treadmill. Whether the melting-asset charge is a manageable cost of goods or an unpayable tax on reported profit depends entirely on two things the balance sheet does not tell you: how fast the revenue compounds, and how long the assets actually last. The rest of this essay is a fight over those two variables.

3. The steelman for rationality

The most careful attempt to reason about whether this spending can pay was published in October 2025 by the podcaster Dwarkesh Patel and his co-author Romeo Dean, in an essay titled Thoughts on the AI buildout. They wrote it, in their own words, "to teach ourselves about the AI buildout," and they modeled two futures rather than betting on one: "explosive growth, and AI winter" [23]. What follows draws the bull-leaning half of that framework into its strongest form; the essay returns below to a premise Patel and Dean share with the bears. Their case deserves to be represented in its own terms before it is engaged, because it is the named live thesis this column was assigned to answer, and because it is more careful than the bull noise around it.

Their central claim, stated verbatim: "In 2025, NVIDIA will turn around $6B in depreciated TSMC Capex value into $200B in revenue" (Tier C, first-party analysis; the authors' own thread) [23]. The move is to trace the value chain to its narrowest point. Taiwan Semiconductor Manufacturing Company, the chip fabricator on which the entire edifice rests, is an interested incumbent with every reason to sound bullish; but its numbers are audited, and they are extraordinary. TSMC turned over roughly $122.5 billion of revenue in 2025, and it has guided its 2026 capital-expenditure program to $52–56 billion (Tier C, compiled analyst estimates) [4]. Patel and Dean's point is that Nvidia, the chip vendor and the most interested party of all in the buildout continuing, sits on top of a sliver of TSMC's depreciated fab capacity and converts it into revenue at a ratio so lopsided that the supply chain beneath it looks radically under-built relative to the value it carries. Their conclusion follows the economist's instinct: "As @tylercowen says, do not underrate the elasticity of supply" [23]. If the returns to compute are this concentrated, capital will find a way to build more fabs, more power, more of everything, and the buildout is the market doing exactly what it should.

To convey how concentrated that value is, Patel and Dean offer a comparison that lands harder than any ratio: a single year of Nvidia's revenue nearly matched the past twenty-five years of combined research, development, and capital expenditure of the largest semiconductor-equipment firms, including ASML, Applied Materials, and Tokyo Electron [23]. Read one way, that is a warning about fragility, a tower of value balanced on a needle of fab capacity. Read the way the authors intend it, it is an argument that the buildout is under-supplied rather than over-supplied — that the returns to compute are so concentrated the rational response is to build far more of the scarce input, not less. That is the strongest form of the bull case: not that AI will change the world, a claim that proves nothing about capital allocation, but that the unit economics of the chokepoint make underinvestment the expensive mistake.

The demand side of their steelman has only strengthened since they wrote. Nvidia (again, the interested party, the vendor whose fortunes rise with every dollar of this spending) reported record revenue of $57.0 billion in its fiscal third quarter of 2026, up 62% year over year, with data-center revenue alone at a record $51.2 billion (Tier A, company earnings release) [6]. Its data-center revenue had already leapt from $15 billion to $115 billion over the two fiscal years ending January 2025 (Tier B, financial reporting) [7]. This is not a company shipping into a warehouse. Downstream, the buyers are monetizing. The research firm Gartner forecasts worldwide end-user spending on AI models and platforms to reach $64 billion in 2026, up 63.4% from $39 billion in 2025, with spending on generative-model software alone forecast to grow 117% (Tier B, analyst forecast, July 2026) [5]. The model labs are scaling into that spend: OpenAI reached roughly $25 billion in annualized revenue by early 2026, up from $13.1 billion in 2025, and Anthropic's run-rate revenue crossed $47 billion by May 2026, up from roughly $9 billion at the end of 2025, with Anthropic projecting its first quarterly operating profit, around $559 million on $10.9 billion of revenue, in the second quarter of 2026 (Tier C, research compilation) [8]. Anthropic books some cloud-reseller revenue gross, which flatters the top line, and single-vendor run-rate figures move fast enough to be treated as directional rather than precise. Even discounted, the direction is unambiguous.

So is the demand under the demand. The neocloud specialists that rent GPUs to the labs are contracted years out: CoreWeave carries a $66.8 billion backlog against 2026 revenue guidance of $12–13 billion, and FluidStack has signed a $50 billion compute offtake agreement with Anthropic, the largest of its kind (Tier C, industry analysis) [9]. When Sundar Pichai told investors that "our AI investments are redefining what's possible across every part of our business" [1], he was making a marketing claim. When his cloud segment grew 82% at a 35-point margin, he was reporting a fact. The steelman, fairly stated, is this: the buildout is a rational option on a scarce and monetizing asset, and the firms that underbuild it will spend the 2030s explaining to shareholders why they ceded the most valuable real estate of the era to bolder rivals.

One caveat travels with their framework. Patel and Dean take the chips themselves to be short-lived, writing that "Every 3 years, the chips depreciate, and the race starts anew" [23]. On the economic life of the silicon, in other words, this column's strongest bull source stands with Michael Burry rather than against him. Their optimism is about the scarce foundry output, not the durability of any given chip on top of it.

4. The counterargument, at full strength

Now the case against, and it must be put at its strongest, because it is strong.

The most damaging attack is not that demand is fake. It is that the reported earnings underneath the buildout are engineered. The investor Michael Burry (who called the 2008 housing collapse, and who disclosed put options on Nvidia and Palantir shortly before making his argument, and is therefore an interested party talking his own book) contends that the hyperscalers are depreciating AI chips over five and six years when the economic life of a graphics processor, given a two-to-three-year product cycle, is far shorter. He estimates the industry will understate depreciation by $176 billion between 2026 and 2028 [10], enough that by 2028 Oracle would be overstating earnings by 26.9% and Meta by 20.8% (Tier C, an interested party's own estimate, since these percentages are Burry's model output rather than audited figures) [27]. This is not a valuation quibble but an accusation that the profit lines investors are underwriting the boom against are inflated by the very schedules that make the assets look durable.

The accounting is real, not hypothetical, and it moves in both directions in ways that reveal how much discretion sits in the estimate. Amazon, reading the pace of AI hardware turnover, shortened the useful life of a subset of its servers from six years to five in early 2025, taking a $700 million hit to operating income and booking $920 million of accelerated depreciation [11]. Meta went the other way in the same window, extending server lives to 5.5 years and thereby reducing depreciation expense by $2.9 billion, close to four percent of its estimated earnings [11]. Google, Oracle, and Microsoft have all pegged AI-server life at up to six years [7]. When a single assumption swings reported profit by billions and the assumption is chosen by the firm being measured, the profit becomes an opinion.

The obsolescence risk that assumption papers over is visible in the rental market, the closest thing to a spot price for the asset. The hourly rate to rent an Nvidia H100 collapsed from more than $8 in 2023 to $1–2 by late 2024 as supply caught up [11]. An asset whose rental yield falls by three-quarters in eighteen months is not obviously a six-year asset.

Layer onto the accounting the financing, and the picture darkens. The spending has outrun the cash flow, so the money is increasingly borrowed. The top five hyperscalers issued $121 billion of bonds in 2025, more than four times their $28 billion five-year average [15]. Technology firms had borrowed roughly $450 billion from private-credit funds by early 2025, with about $125 billion flowing into data-center project finance that year alone, and the debt is migrating off the balance sheet entirely: Meta raised roughly $30 billion for a single Louisiana facility through a special-purpose vehicle, the Beignet fund, backed by private-credit lenders [15, 16]. That machinery is not universal: the SPV structures cluster around Meta, Oracle, xAI, and the neoclouds, while Google, Microsoft, and Amazon have so far funded their data centers with cash and conventional bonds [15, 16]. OpenAI, the demand engine at the center of it all, has committed to more than $1.4 trillion of long-term compute contracts against a fraction of that in revenue [16]. The strain is already visible on the most exposed balance sheet: Oracle carries about $131.7 billion of total debt against $17.7 billion of operating income, a ratio of 7.4 times, and it is one of the firms Burry singles out for the depreciation adjustment that would most flatter its reported profit [15]. A company this levered, monetizing a buildout this early, is precisely where a demand disappointment and a credit-market tightening would meet.

And the capital does not merely flow outward; it circles. In September 2025 Nvidia (the chip vendor, financing its own customers) announced an investment of up to $100 billion in OpenAI, which contracts with Oracle for some $300 billion of cloud over five years, with Oracle spending about $40 billion on Nvidia's own GB200 chips to fulfill it [13, 14]. Money leaves Nvidia, becomes OpenAI's spending, becomes Oracle's capex, and returns to Nvidia as revenue. Critics who lived through the telecom-vendor-financing collapse of 2001 recognize the shape: a closed loop in which the same dollar is counted as investment, spending, and sales on three different income statements, inflating all of them.

The bear's closing move is to ask what any of this has produced for the customers footing the ultimate bill. A 2025 study by MIT's NANDA initiative, "The GenAI Divide," examined 300 enterprise deployments and found that 95% delivered no measurable profit-and-loss impact, with only 5% generating significant value [17]. If the enterprises paying for AI cannot show a return, then the revenue growth financing the buildout rests on subsidized experimentation, venture capital, and the labs' own losses — a base that thins the moment capital gets expensive. Put the accusations together and the bear case is not that AI is useless; it is that a genuine technology has been financed like a fraud, on inflated earnings, borrowed money, and circular revenue, and that the correction would be violent because the most aggressive spenders are levered to the same three assumptions.

5. Where the counterargument holds, and where it breaks

That case is coherent, and on two of its three legs it is right. It is the third leg, obsolescence, that gives way under weight, and once it does, the whole edifice of the bubble thesis reorganizes into something narrower and more precise.

Take the depreciation-obsolescence claim on its own terms. Burry's premise is that a chip on a two-to-three-year product cycle cannot be a five-or-six-year revenue asset. The premise confuses the release cadence of new chips with the earning life of old ones, and the market has been running a live experiment on exactly this question. Google's own engineers report that seven- and eight-year-old tensor processing units run at full utilization, with seven hardware generations in production simultaneously [11]. CoreWeave's Nvidia A100 chips, announced back in 2020, remain fully booked, and when a batch of H100 contracts expired the units were immediately re-leased at 95% of their original price [7]. Most telling, the rental collapse the bears point to has reversed: the SemiAnalysis one-year H100 contract index (from the analyst house SemiAnalysis, a bullish and interested observer, but one drawing on survey data from more than a hundred market participants) rose from $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, a roughly 40% increase, with some H100 contracts now signed four years out through 2028 [12].

An asset re-leasing at 95% of its original price years after purchase, with contract prices rising, is behaving like scarce, durable infrastructure rather than like a melting ice cube — for now. The reason is physical, and it is the same reason the whole buildout is bottlenecked, which I will come to: there are not enough chips, not enough memory, and not enough power to render the installed base obsolete on schedule. Obsolescence requires a replacement to exist and be available. When the replacement is itself sold out, last year's silicon keeps earning.

So the honest reading is narrower than the bears' and narrower than a triumphant bull's: the obsolescence timing is looser than Burry assumes, conditional on the shortage holding. That condition is not free, and it sits in tension with the bull case's own engine. If the elasticity of supply that Patel and Dean invoke actually delivers the fabs, the power, and the next chip generation, then the replacement arrives and obsolescence returns on something closer to the bears' clock, the very three-year cadence Patel and Dean themselves assume. The durability is real, then, but rented from the shortage rather than owned outright.

But notice what survives this rebuttal, because it is precisely the part the bull case would prefer to bury. The obsolescence leg was the one that justified the accounting attack — and it is the only leg I have knocked out. The accounting discretion is still real: Meta's $2.9 billion swing and Amazon's opposite adjustment prove that reported AI profit is a choosable number, and a firm choosing a six-year life is choosing to report earnings sooner and smoother than a five-year life would allow [11]. The circular financing is still real: the Nvidia–OpenAI–Oracle loop counts one economic event as three, and no rebuttal about GPU longevity makes a dollar counted three times worth three dollars [13, 14]. The leverage is still real, though unevenly distributed: $121 billion of bonds at four times the historical rate, and tens of billions moved into special-purpose vehicles, convert the most aggressive spenders from firms that used to self-fund into ones that depend on the continued generosity of credit markets [15, 16].

Here, then, is where the argument lands. The bears are right that a bubble-shaped risk exists, and wrong about where it lives. It does not live in the asset but in the paper wrapped around it. The compute is durable enough to be worth owning; the six-year schedules and the round-tripped, off-balance-sheet financing are the machinery that promises the returns will arrive faster, smoother, and more certainly than a genuinely scarce and slowly monetizing physical asset can honestly deliver. Strip the engineering away and you are left with a rational, patient, capital-intensive infrastructure bet. Leave the engineering on and you have a rational bet levered and marked in a way that turns an ordinary cyclical disappointment into a potential cascade. The bubble is not the buildout. The bubble is the leverage and the ledger.

6. The physical floor under the paper

The reason the asset holds its value — and the reason the buildout cannot inflate as fast or as recklessly as pure financial mania would — is that it runs into physics long before it runs into imagination. This is the substrate beneath the whole debate, and it disciplines both the bull's elasticity story and the bear's overbuild story.

Power is the first wall. The International Energy Agency estimates that data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of the world's total, and projects that to more than double to about 945 terawatt-hours by 2030, driven mainly by AI (Tier A, IEA data via official EU publication) [18], a level just under 3% of global electricity demand [19]. In the United States the increase is sharper, with data-center electricity demand projected to grow about 130% by 2030, and in China about 170% over the same horizon (Tier A, IEA release) [19]. That demand does not materialize because a spreadsheet wills it. It waits on interconnection queues, transformers, and turbines. GE Vernova, one of the few firms that can build the gas generation the grid needs, saw its equipment backlog swell from 44 to 53 gigawatts in a single stretch, with reservation agreements climbing from 56 to 63 gigawatts (Tier B, company results) [22]. When the power equipment is reserved years in advance, the buildout's ceiling is set by the turbine line, not the capital markets.

Memory is the second wall. High-bandwidth memory, the component that feeds an AI accelerator, has been in shortage for over a year, and the manufacturers expect it to persist: Samsung's memory chief warned in its April 30, 2026 earnings report of significant shortages across memory products continuing through at least 2027 [20]. Memory prices were expected to rise more than 50% in the first quarter of 2026 alone [21]. Fab capacity is the third wall, the one TSMC's $52–56 billion capex program is straining to push back, and even that will lift its depreciation by a high-teens percentage in 2026, pressuring the very margins the buildout is supposed to generate [24].

These constraints cut against the bulls and the bears at once, and on the bull side the sharpest statement of them belongs to Patel and Dean themselves. It was they who argued that lead times are "the overwhelming consideration" in how a data center is powered, that, in their words, "Every month that the shell is not set up is a month that the chips ... aren't being used," and that the explosive-growth path would drive "global AI power draw ... twice US's current electricity generation" [23]. Their elasticity-of-supply optimism arrives already fenced by physics: you cannot conjure a gigawatt of power or a fab node on the timeline that spreadsheet returns assume. Against the bears, the same walls are what keep the installed base valuable: a chip cannot be made obsolete by a successor that cannot be powered, cooled, or fed with memory. Physical scarcity is the floor under the asset and the brake on the mania. It is the reason this is a bet worth making and a bet that cannot be made as fast as the capital wants to make it.

7. The load-bearing uncertainty

Everything above resolves into a single question that the evidence cannot yet close, and honesty requires naming it plainly rather than burying it in qualifications.

The bet is rational if, and only if, revenue compounds faster than the combined drag of the three-year chip-replacement clock and the interest on the borrowed money. The asset is durable, but it is not permanent; the financing is available, but it is not free. Between the demand data and the doubt data sits a genuine tension that no amount of rhetorical confidence dissolves. On one side, Anthropic's projected swing to an operating profit and Google Cloud compounding at 82% suggest the revenue is not merely real but accelerating past the cost curve [1, 8]. On the other, MIT's finding that 95% of enterprise deployments show no measurable return suggests the ultimate customers have not yet found the productivity that would make the labs' revenue durable rather than experimental [17]. Both can be true today. They cannot both be true in 2028. Either the enterprise return arrives and validates the run-rates, or it does not and the run-rates prove to have been venture capital and vendor loops wearing the costume of demand.

I have taken my position — that the asset is sound and the financing is where the danger sits — and I hold it. But the position rests on the wager that demand growth wins the race against depreciation and debt service, and that wager is not yet settled. This is the one place the essay concedes uncertainty, and it is the load-bearing one.

So let me make the position falsifiable, because a thesis that cannot be proven wrong is not a thesis. My claim fails, and the bears win outright, if by the fourth quarter of 2027 the SemiAnalysis H100 contract index has fallen back below its October 2025 level of $1.70 per GPU-hour on sustained slack rather than shortage, indicating the installed base is genuinely being made obsolete. It fails if a major hyperscaler is forced to take a large impairment on AI infrastructure and simultaneously extends depreciation lives further to cushion the hit, confirming Burry's engineered-earnings thesis in the same filing. And it fails if enterprise adoption studies through 2027 continue to show the measurable-return rate stuck near the 2025 figure of 5% [17]. Conversely, the bear case fails if the labs' aggregate revenue clears the depreciation-and-interest hurdle on unsubsidized enterprise spending before the next chip generation resets the clock. And the marker that would confirm this essay's particular frame — that the danger is financial rather than fundamental — is a break that originates in the credit stack rather than in the order book: a neocloud or a data-center special-purpose vehicle forced to restructure its private debt while GPU utilization stays high and rental prices hold. Demand intact, financing broken, would be the signature of the failure I think most plausible. Watch those markers. They will settle this before the rhetoric does.

8. What it means for builders, investors, and operators

The distinction between a rational asset and a fragile ledger is not a debating point; it changes what each actor in this economy should do next.

For builders, the labs and the platforms doing the spending, the lesson is that the asset will forgive patience but the financing will not. Owning scarce compute is defensible; the durability data says so. Financing it with structures that assume a smooth, early return is where builders convert a survivable downturn into an existential one. The firm that funds its buildout from cash flow and honest depreciation schedules will still be standing when credit tightens; the firm that has moved its obligations into special-purpose vehicles and stretched its asset lives to flatter its earnings has borrowed against a future it does not control. The competitive pressure to underreport depreciation is real, and it is a trap: the earnings it manufactures today are the impairments it discloses tomorrow.

For investors, the actionable move is to stop reading the income statement and start reading the notes. Reported AI profitability is, as Meta and Amazon demonstrated in opposite directions, a chosen number; the useful-life assumption and the depreciation schedule are where the real economics hide, and two firms with identical hardware can report profits that differ by billions on that choice alone [11]. The revenue growth is the bull signal and it is largely trustworthy; the earnings quality is the bear signal and it requires forensic attention. An investor who owns the picks-and-shovels scarcity — power, memory, fabs — is buying the floor. An investor who owns the leveraged, off-balance-sheet financing of it is buying the fragility. They are not the same trade, and the market has spent 2026 pricing them as though they were.

For operators, the utilities, grid planners, memory makers, and turbine builders, this is the least ambiguous position in the entire economy, and the most durable. The bottleneck is the business. The IEA's doubling of data-center power demand, the memory shortage stretching through 2027, the multi-year turbine backlogs: these are not risks to the operator, they are the demand curve [18, 20, 22]. The operator's exposure is not that the AI boom is a bubble; it is that the operator underbuilds the physical capacity and cedes a decade of pricing power. Where the builder must fear the financing and the investor must fear the ledger, the operator's only real error is timidity.

And standing behind all three now is the state. On November 24–25, 2025, the administration launched the Genesis Mission by executive order, directing the Department of Energy to build a national AI platform for science, and by March 2026 the DOE had opened a $293 million funding call to seed it (Tier B, official reporting) [25, 26]. The dollar figures are small beside the hyperscalers' hundreds of billions, and the original order committed no funding at all. But the signal is not the money; it is that the government has declared the buildout strategic infrastructure, which changes the downside: an asset the state considers a matter of national security is an asset the state is unlikely to let fail cleanly. That is a floor under the compute and, less comfortably, a subsidy to the very financing structures that most deserve scrutiny.

The buildout, in the end, is neither the triumphant certainty its boosters sell nor the fraud its shorts are positioned for. It is a rational bet on a genuinely scarce thing, made by disciplined firms who understand the asset better than their critics do, and then dressed in financial engineering that borrows against a confidence the physics has not yet earned. The chips will keep earning. The question that will define the next three years is whether the paper stacked on top of them can survive the moment the market stops taking the depreciation table on faith. Bet on the compute. Read the notes.