The $650 Billion AI Race No One Can Quit

In early July, Meta launched a service called Meta Compute to sell excess GPU capacity to other companies. The product was unremarkable. What Mark Zuckerberg said to justify it was not.

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

This is the logic of the AI build-out spoken aloud. Keep spending, because if you overbuild, you can always sell the surplus. The problem is that this only works if someone else underbuilt. When every hyperscaler is racing to spend more, there is no one left to buy. CoreWeave and Nebius, two companies whose entire business is selling GPU compute, saw their stocks drop 17% on the news ^[1]. The market understood what Zuckerberg had just admitted: the capacity glut is already arriving. The spending numbers have become difficult to hold in your head. Amazon projects roughly $200 billion in capex this year. Alphabet raised its guidance to $180 to $190 billion. Meta set a floor of $125 billion. Collectively, the four largest hyperscalers plan to spend $650 billion on AI infrastructure in 2026 alone ^[2][3]. Nvidia, which sits at the center of this as the dominant GPU supplier, projects hyperscaler capex will hit $1 trillion in 2027 and $3 to $4 trillion annually by 2030 ^[4]. The market has been telling these companies to stop. Amazon shares fell 11.5% after it announced its $200 billion plan. Meta dropped 12%. Microsoft is down 20% on the year. Oracle has been cut nearly in half ^[5][6]. Not one of them has flinched. Amazon CEO Andy Jassy made the logic explicit.

We believe that AI is the biggest technology transformation in our lifetimes. — Andy Jassy

He is not wrong about the opportunity. He is describing the trap. Any company that cuts spending cedes the race, so all of them keep spending, even as the collective bill grows beyond what any plausible revenue trajectory can justify. The bill is being paid with borrowed money, and the borrowing is reaching for new forms because the old ones are straining. Amazon executed the largest corporate bond sale in history in March, raising $53.8 billion ^[7]. Alphabet issued a 100-year "century bond" in February, the first by a technology company since Motorola in 1997. Analyst Bill Blain's reaction was widely quoted.

I think the fact that a 100-year bond comes out, you can't get much more frothy than that. — Bill Blain

Nvidia raised $25 billion in its largest-ever bond sale despite sitting on $50 billion in cash ^[8]. CoreWeave secured an $8.5 billion loan collateralized by its GPU hardware, a structure that did not exist before this cycle ^[9]. Morgan Stanley forecasts hyperscalers will borrow $400 billion in 2026 to fund the build-out ^[10]. Then there is the debt that does not appear on any balance sheet. A Nikkei investigation published this week found that five tech giants hold $1.65 trillion in off-balance-sheet AI obligations, exceeding the $1.35 trillion they report as conventional debt ^[11]. Meta alone carries roughly $420 billion in hidden liabilities, including a joint venture with Blue Owl Capital for its Hyperion data center. Oracle's off-balance-sheet commitments grew thirtyfold in four years to support OpenAI's compute needs. The funding mechanism is showing fatigue even as the spending accelerates. When Amazon returned to the bond market in July for another $25 billion, it had to offer 18 to 21 basis points of extra yield to attract buyers. Bank of America's read on the situation was blunt.

once-in-a-lifetime opportunity. — Andy Jassy

Amazon informed underwriters this would be its final debt issuance for 2026 ^[12]. The free cash flow that once made these companies the safest credits in the world is evaporating. Amazon's fell from $38.2 billion to $11.2 billion year over year ^[6]. Meta reported $43.6 billion in free cash flow for 2025, but $42 billion of it was consumed by stock-based compensation costs, leaving real liquidity near zero ^[13]. Reuters reported this week that US hyperscalers are on track to spend more on combined capex than they generate in free cash flow by 2027 ^[14]. Debt must be serviced by revenue, and the revenue model is under attack from three directions at once. The first is price. In May, DeepSeek permanently cut the price of its V4-Pro model by 75%, making it 12 to 19 times cheaper than OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 for equivalent tasks ^[15]. Analyst Sanchit Vir Gogia called it an efficiency gain being passed through rather than a promotion ^[15]. When the marginal cost of intelligence collapses, the per-unit revenue that was supposed to amortize $650 billion in annual infrastructure spending collapses with it. The second is the customer. PNC Financial CEO Bill Demchak stated the problem in terms any CFO would understand.

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

PNC is building its own GPU compute to avoid paying external token costs ^[16]. It is not alone. Companies are shifting to smaller, cheaper models and open-source alternatives as hyperscalers transition from flat subscriptions to usage-based pricing. The customers who were supposed to generate the return on all this infrastructure are finding ways to need less of it. The third is the hyperscalers themselves. Microsoft cancelled thousands of internal Claude Code licenses and redirected engineers to GitHub Copilot CLI after discovering that compute costs exceeded human labor costs. Uber exhausted its entire 2026 AI tool budget in four months. Nvidia VP Bryan Catanzaro, whose own company sells the chips powering this build-out, said it plainly.

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

The companies spending $650 billion on AI infrastructure are simultaneously laying off 92,000 workers and discovering that the AI tools they are building are too expensive for their own teams to use ^[17]. This is where the bulls' favorite analogy breaks down. The argument, repeated by Morgan Stanley, Goldman Sachs, and Empower Investments as they urge clients to buy the dip after the Magnificent Seven lost $2.2 trillion in market cap in June, is that this cycle mirrors the 2016 to 2020 cloud build-out ^[18]. The cloud era also saw massive upfront spending, market skepticism, and eventual dominance. The playbook worked. But the differences are structural. The cloud build-out was perhaps a tenth the scale of what is underway now. There was no price competitor offering a functionally equivalent product at a 75% discount: AWS was not competing against a Chinese cloud provider selling compute at one-nineteenth the price while it was still building data centers. The debt was on the balance sheet, not hidden in joint ventures and GPU-collateralized loans. And the revenue model was not under attack from the customers, the competitors, and the builders themselves all at once. The counter-evidence is real and should be stated plainly. Revenue is growing. Alphabet's cloud division grew 63%. Meta's ad revenue rose 33% year over year to $56.3 billion in the first quarter. Microsoft reported $82.9 billion in revenue with a $37 billion AI annual revenue run rate and Azure growing 40% ^[19]. Micron's high-bandwidth memory is sold out through 2028 under non-cancellable contracts ^[20]. The physical supply chain shows no sign of demand contraction. The question is not whether revenue is growing. It is whether revenue can grow fast enough to outrun the carrying cost of $650 billion in annual capex, funded by increasingly strained debt, when three forces are simultaneously compressing the price at which that revenue can be earned. The BIS put the risk in historical terms in its June annual report, comparing the current boom to the dot-com bubble and the 1830s canal mania and warning that such episodes ended with investment reversals that induced economy-wide recessions ^[21]. A US Treasury draft report reached similar conclusions before the administration dismissed it as unvetted ^[22]. Senator Elizabeth Warren proposed legislation requiring AI companies to disclose their off-balance-sheet debt ^[22]. The prisoner's dilemma has a third arm that the cloud-era analogy misses entirely. You cannot stop spending, because your rivals will not. You cannot collectively stop, because no coordination mechanism exists. And you cannot control the price at which you sell what you have built, because DeepSeek, your own customers, and your own cost-conscious engineering teams are all pushing it down. In the cloud build-out, the revenue model held steady while the infrastructure went up. This time, the revenue model is eroding at the same moment the bill is coming due.