Everyone finances everyone else's computeEveryone finances everyone else's computeEveryone finances everyone else's computeEveryone finances everyone else's computeEveryone finances everyone else's computeEveryone finances everyone else's computeEveryone finances everyone else's compute

The five largest infrastructure buyers are on track to spend somewhere in the region of $700–800 billion on AI infrastructure this year. Goldman’s projection has hyperscaler capex at roughly $1.15 trillion across 2025–2027, against $477 billion for the whole 2022–2024 stretch. Whatever this is, it is not incremental.

Two structural details make it more interesting than a large number.

The circle

A recurring pattern in the deal flow: a chip vendor takes equity in a model lab, the lab signs take-or-pay compute commitments with a cloud provider, the cloud provider funds GPU purchases with debt, and the GPUs come from the chip vendor. Nvidia, OpenAI, Oracle and CoreWeave appear in each other’s arrangements in interlocking positions, and capex has outrun operating cash flow far enough that hyperscalers are reaching for external financing.

None of that is illegal or even unusual in a capital-intensive buildout. What it does is make end demand look larger and more independent than it is. A dollar can be counted at several points around the loop. The headline Nvidia–OpenAI figure spent most of a year as a non-binding letter of intent before resolving into a signed guarantee — a distinction that matters enormously and gets flattened in every summary.

The depreciation question

The one that will actually settle the argument:

• Hyperscalers generally depreciate AI hardware over four to six years, and compute is heading toward roughly 60% of capex, up from around 43% in 2022 — so the assumption matters more each quarter.
• Michael Burry has argued the real economic life is closer to two to three years, tracking Nvidia’s three-year generational cadence, and estimated the cumulative hit to reported earnings at over $176 billion across 2026–2028 if the schedules were shortened to match.

Attribute that figure carefully: it is one prominent investor’s estimate under his own assumptions, not a consensus. But the disagreement is not confined to short sellers — it is live inside the industry. In January 2025 Amazon shortened the useful life of part of its fleet and retired hardware early, citing the pace of AI hardware development. The same month, Meta extended its schedule to 5.5 years.

Two of the most sophisticated operators in the world, looking at the same technology cycle, moved their assumptions in opposite directions. That is what an unsettled question looks like on a balance sheet, and it is load-bearing under every “the capex is fine” argument being made.

For anyone building on top

The exposure is not to the stock price. It is to the price of tokens, which today is set by companies willing to lose money to hold a position, and which is therefore not a real price. Build as if inference costs what it costs to serve, not what it costs to buy.

If your architecture only works at subsidised prices, you have not built a business. You have taken a position on somebody else’s balance sheet.