The Numbers Come First
Apollo Chief Economist Torsten Slok published a blog post Friday laying out a structural problem at the heart of the AI boom: the companies closest to the end customer are the ones losing the most money.
Using data from Pitchbook and Bloomberg, Slok divided the AI value chain into four segments — models and applications, cloud and compute, energy and grid, and silicon and equipment. His finding is stark. Chipmakers and equipment suppliers, the furthest upstream, carry a 41% profit margin. Models and applications companies — the layer that includes firms such as Anthropic — run at a -59% operating margin.
The implication is direct. 'AI boom's profits are currently being funded by investors rather than earned from customers,' Slok wrote. 'The upstream margins are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand.'
A Chain Only as Strong as Its Weakest Link
Slok's warning is not that AI fails as a technology. It is that the financial architecture holding the boom together is fragile. 'The most profitable part of the AI value chain depends on the least profitable part continuing to grow revenue or raise capital,' he concluded. 'Capital can bridge the gap for a while, but not indefinitely.'
Goldman Sachs now projects AI investment will exceed $1 trillion in 2026. Yet outside the Magnificent Seven, there is no significant change in economic productivity or profit margin growth to justify that figure, according to Slok.
The Bank of International Settlements flagged the same fault line in its annual report published in June. The five major hyperscalers are issuing debt to fund capital expenditures that are outpacing earnings and free cash flow. A Bank of America analysis from last November found those five hyperscalers issued $121 billion in debt in 2025 — four times their average annual debt issuance over the prior five years.
'Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust,' the BIS report warned.
Oracle as the Stress Test
Tech writer Ed Zitron pointed to Oracle as the clearest example of the exposure. As of the end of fiscal 2026, Oracle carries a negative cash flow of $23.7 billion, nearly $130 billion in outstanding debt, and $260 billion in lease commitments for AI infrastructure projects that have not yet begun. That buildout is tied directly to a $300 billion compute deal signed with OpenAI last September.
'Oracle's existence — and Larry Ellison's personal wealth — hinges on whether OpenAI can make good on its promise to spend $300bn in compute,' Zitron wrote in a Substack post in June.
What the Market Has Not Priced
The numbers here are not a call to dismiss AI as a technology. They are a call to read the balance sheet honestly. When the most profitable segment of an industry depends entirely on the least profitable segment continuing to attract outside capital, the system is not self-sustaining — it is subsidized.
Free markets reward genuine productivity gains. They eventually punish structures where upstream margins exist only because downstream losses are being financed by investors chasing a narrative. The question Slok poses — will ROI show up for AI's end customers fast enough to sustain the spending generating those upstream margins — is the only question that matters now. Capital is patient, but it is not infinite, and the bond market will price the answer before the press release arrives.
