One of Wall Street’s most closely followed economists has put a blunt frame on the artificial intelligence boom: the money being made in AI is not, for the most part, coming from customers.
Torsten Slok, chief economist at Apollo Global Management, argues that the sector’s profits “are currently being funded by investors rather than earned from customers” — a distinction that separates a durable business from a financing cycle, and one that becomes uncomfortable the moment the financing slows.
The margin split that drives the argument
The clearest evidence Slok points to is the gap between the two halves of the AI supply chain.
The silicon and equipment companies — the chipmakers and infrastructure suppliers selling the picks and shovels — are running at roughly a 41% operating margin. The model and application companies buying from them are running at approximately negative 59%. Anthropic is cited as an example on the losing side of that ledger.
That is not a subtle spread. It means value is flowing decisively in one direction: from companies burning capital raised in private and public markets, into companies selling them hardware. As long as the capital keeps arriving, the chipmakers’ revenue looks like customer demand. If it stops, a large share of that revenue turns out to have been someone’s fundraising round rather than someone’s operating budget.
The scale of the bet
The numbers around AI capital expenditure have grown large enough to move macroeconomic aggregates. Goldman Sachs projects AI investment will exceed $1 trillion in 2026.
More telling is how it is being paid for. The Big Five hyperscalers issued $121 billion in debt during 2025 — roughly four times their average annual debt issuance over the previous five years. These are companies historically defined by enormous free cash flow and minimal need for external financing. A fourfold jump in borrowing is a structural change in how the buildout is funded, not a rounding error.
Debt-financed capex is not inherently reckless; utilities and telecoms have built infrastructure that way for a century. But it does convert a flexible cost into a fixed obligation. Cloud capex funded from operating cash can be dialled back in a bad quarter. Interest payments cannot.
Oracle as the stress test
Oracle is where the argument becomes concrete, and Slok’s case leans on it heavily.
The company ended fiscal 2026 with negative cash flow of $23.7 billion and carries nearly $130 billion in outstanding debt. Against that, it has committed to roughly $260 billion in AI infrastructure leases — obligations that must be serviced regardless of how much AI revenue materialises — and in September 2025 announced a $300 billion deal with OpenAI.
The structure of that arrangement is what draws scrutiny. A $300 billion commitment from a company that is itself deeply lossmaking and dependent on continued fundraising is a receivable whose quality rests on the funding environment holding up. Oracle has borrowed and leased against that expectation. If OpenAI’s capital access tightens, Oracle’s lease obligations do not shrink to match.
Who else is warning
Slok is not alone. The Bank for International Settlements used its annual report in June 2026 to warn that a pullback in financing could trigger an investment bust — the standard mechanism by which a capex cycle turns, in which the marginal lender’s caution becomes the whole market’s contraction. Bank of America has weighed in on the financing structures involved. The critic Ed Zitron has made the narrower point that the entire edifice depends on a handful of major technology companies continuing to spend, and that any one of them stepping back changes the arithmetic for everyone selling into them.
Nvidia, AMD and Micron sit on the profitable side of the split; Microsoft, Amazon and Constellation Energy are all named in the buildout, the last of them because power supply has become one of the binding constraints on data center expansion.
What the bears might be missing
The counterargument deserves a fair hearing. Negative operating margins at model companies are partly a definitional artifact: training runs are expensed as incurred while the resulting model generates revenue over subsequent years, so the accounting understates the economics of a business in heavy build phase. Amazon posted thin or negative margins for years while building infrastructure that later proved enormously valuable.
The distinction that matters is whether spending is building an asset with a long useful life or renting a depreciating one. GPUs are not railways. They lose competitive relevance in a few years as successive generations arrive, which makes the payback window much shorter than the infrastructure analogies suggest.
What to watch
Three indicators will show whether Slok is right. Whether enterprise AI revenue — money paid by customers for delivered software, not by investors for equity — grows fast enough to close the margin gap. Whether hyperscaler debt issuance continues at 2025’s elevated pace or normalises. And whether any large buyer publicly trims its capex guidance, which historically is the moment a capex cycle turns from confident to defensive.
The broader corporate earnings picture has already shown how much reported profitability can depend on factors unrelated to customer demand, as with Sony’s 37% profit jump on tariff refunds. The physical footprint of the buildout is drawing its own scrutiny, including Amazon’s Texas data center and its emissions profile.
