The AI Capex Machine: When Does It Stop?

In the critically acclaimed movie Margin Call, on the eve of the subprime market crash, the head of the fictional investment bank John Tuld expressed the weight on his shoulders: I'm here for one reason and one reason alone. I'm here to guess what the music might do a week, a month, a year from now. That's it. Nothing more. And standing here tonight, I'm afraid that I don't hear — a — thing. Just… silence. Make no mistake, we are not here to herald an incoming market crash — far from it. We are borrowing Tuld's analogy to listen to the market today, and to guess what it may sound like a few months or a year from now. And for now, the music is loud — played overwhelmingly on the note of AI capex spending, if not solely.

The numbers are gargantuan

The four largest AI infrastructure investors — Microsoft, Amazon, Alphabet, and Meta — reported their latest earnings within hours of each other on April 29, 2026. Combined, they have committed approximately $700 billion in capital expenditure for the calendar year, up 77% from 2025. To put that in perspective, this exceeds the GDP of Switzerland and is roughly equivalent to the entire annual defence budget of the United States. Every quarter, these companies are building the equivalent of a medium-sized country's infrastructure — in data centres, GPUs, and cooling systems — betting that artificial intelligence will generate returns large enough to justify it.

The question is no longer whether the AI business is real. It is. Cloud revenue is accelerating. Azure grew 40% year-on-year, Google Cloud grew 63%, and AWS posted its fastest growth in fifteen quarters. Enterprises are adopting AI tools across every industry. Utilisation rates at major data centres run above 80%. This is not a speculative bubble built on vapour — the technology works, and people are paying for it.

The question is whether the spending can sustain itself.

The math that rules

Capital expenditure is funded from operating cash flow. When capex exceeds cash flow, a company must either draw down its reserves or borrow. Neither is sustainable indefinitely. We built a model tracking this constraint across the four hyperscalers, starting with Q4 2025 data and updating it with the latest quarter. The results are striking.

Amazon now consumes approximately 99% of its operating cash flow on capital expenditure. Its trailing twelve-month free cash flow collapsed from $25.9 billion to $1.2 billion — a 95% decline — in a single year. To keep building, the company raised $54 billion in new debt in March 2026. The mechanism our model predicted — capex exceeding cash flow, triggering debt issuance — is no longer a projection. It is happening.

Alphabet's capex grew 107% year-on-year, the fastest of the four, and management guided for a “significant increase” in 2027. Their capex-to-cash-flow ratio jumped from 53% to 78% in one quarter. The buffer is eroding fast. Microsoft improved from 84% to 68% on stronger-than-expected cash flow growth, but guided next quarter's capex above $40 billion. Meta raised its full-year guidance to $125–145 billion, and the market immediately wiped $175 billion from its valuation.

Across all four, capital expenditure is growing two to five times faster than revenue. Microsoft spends $2.70 in capex for every $1 of revenue growth. Alphabet spends $4.90. Amazon spends $4.50. Only Meta, at $1.60, is approaching a sustainable ratio — and even that triggered a market sell-off.

Outside the four, Oracle presents a starker picture. With 2 to 3 percent cloud market share, Oracle spent $18.6 billion in capex against $17.2 billion in revenue last quarter — a ratio of 108 percent, funded almost entirely by debt. Its trailing free cash flow is negative $24.7 billion.

AI revenue growth has to accelerate

For the current spending trajectory to be justified, AI revenue would need to accelerate from its current growth rates to match capex growth — roughly tripling in pace. No technology product in history has achieved that kind of acceleration at this revenue scale. The closest precedent is cloud computing itself, which took over a decade to grow from experimental to dominant. AI may get there, but the capex is being deployed on a three-to-five-year payback assumption, not a ten-year one.

There is a deeper structural concern. A meaningful portion of reported AI revenue is circular. Hyperscalers invest in AI startups, who then purchase cloud compute from the same hyperscalers, which gets reported as AI revenue growth, which justifies more capex. Microsoft invested in OpenAI, whose compute bill is Microsoft's Azure revenue. Amazon and Google invested in Anthropic, whose cloud spend flows back to AWS and GCP. When venture funding tightens — and it is tightening as rates remain elevated and AI exit markets stay closed — these captive customers lose their funding, and the reported revenue growth overstates real end-user demand by an unknown but potentially significant margin.

What to watch

The catalyst for repricing is not a technological failure. It is a capital allocation decision. The moment the first hyperscaler guides capex flat or down, the AI supply chain — GPU manufacturers, data centre operators, memory suppliers, and the constellation of companies that exist to service this build-out — reprices. This is a quarterly earnings event, not a daily market signal.

Oracle, the most leveraged player in the build-out, may crack before any hyperscaler — but whether the market reads that as company-specific or systemic depends on timing. Among the hyperscalers, Amazon is the most likely first mover. It is already at the cash flow constraint, already issuing debt, and has the weakest balance sheet of the four. The window is Q4 2026 through H1 2027. The others follow in sequence: Alphabet when its buffer erodes to the 90% threshold, Microsoft when its next capex step-up pushes it back toward cash flow parity, and Meta — which cut Reality Labs spending in 2023 under shareholder pressure — when the board decides enough is enough.

None of this requires AI to fail. It requires AI to be slightly less transformative, adopted slightly more slowly, than a $700 billion annual wager implies. Those are not generous odds.

Note on data

All financial figures are sourced from Q1 2026 earnings releases (April 29, 2026) and company investor relations filings. Oracle figures are from its fiscal Q3 FY2026 earnings (March 10, 2026). Capex figures include property and equipment purchases. Operating cash flow and free cash flow are as reported or derived from reported figures. Forward projections assume current growth rates continue; actual outcomes will differ. This article does not constitute investment advice.

Updated 7 May 2026: Added Oracle capex and balance sheet data.