Does the AI money add up?

Our read

The build is not three or four companies. It is about eight serious builders, and the top four fund most of it from businesses that earned money long before AI: ads, cloud, software and retail.

But three of the four now spend more than their businesses bring in. Free cash flow, the cushion that made this safe, is negative at Amazon, Alphabet and Meta. Microsoft is the exception on paper, but a change in how it counts leases now keeps a growing share of its building out of capex, and it just sold $25 billion of bonds anyway. The gap is being filled with debt.

This is a US story. China's big tech companies spend about a sixth as much, and they too have moved to negative free cash flow and new share sales to keep up.

Our read

This is the narrow part. Outside the big four's own products, the paying demand for new AI capacity is mostly two labs, plus Meta renting overflow. Each lab's signed purchases are roughly twenty times those of the next buyer on the list, Meta.

The Microsoft data point matters and cuts the other way: new backlog from ordinary enterprise customers is real and growing. It is not yet big enough to replace a lab that stops paying.

Our read

In 2024 this was a cash story. In 2026 it is a cash-plus-debt story, and the debt is growing fastest away from the big four, at borrowers with weaker credit. That is where a downturn would bite.

The healthiest sign is who still writes equity checks: public stock buyers for SpaceX and Alphabet, and the biggest growth investors for Anthropic. The least healthy is SoftBank arranging loans against the very shares it is buying.

Our read

Anthropic and OpenAI are not the same risk. Anthropic is roughly self-funding today, holds $132 billion of cash and short-term investments, and is about to tap public markets. Its risk is that its $498B of non-cancellable contracts assumes growth continues for years.

OpenAI is the hinge of the whole system. Its own projections have it spending about as much on compute as it earns for five years, so every dollar of its commitments to Oracle, Microsoft, Amazon and CoreWeave depends on the next funding round arriving. So far each round has, and from bigger names.

Our read

The chain to watch is: lenders fund developers, developers build for Oracle, Oracle rents to OpenAI, and OpenAI pays with money raised from SoftBank, Nvidia and Amazon. SoftBank in turn borrows to pay. Each link is a company with less cushion than the one it serves.

Nvidia is the quiet exposure. It is now lending its balance sheet, through stakes, capacity backstops and lease guarantees, to keep its customers buying. That works while chips are scarce. It turns into losses on Nvidia's books if a big customer stops paying, and a downturn would hit its chip sales, its investment values and its guarantees at the same time. The guarantees start in 2028, so this is a risk for the next cycle, not this one.

Two exposures this page names but cannot yet size: how fast the chips wear out, which decides whether lenders' collateral holds its value; and who ends up owning the power plants if a data-center tenant walks away.

Our read

Lots of companies pay a little, and a few pay a lot. That is the pattern of a technology in its adoption phase, not a bubble with no customers. The median hides the heavy users: the top tenth spend fifty times the typical firm, and consumer subscriptions and AI built into products the firm already buys are not counted here at all. Even so, ordinary business spending does not yet come close to paying for $700 billion a year of data centers. For now the labs, and the big four's own products, carry the load.

Our read

OpenAI cannot get through 2028 on the money it has, even if everything goes to plan. On its own numbers it needs about $60 to 105 billion more by the end of 2028 and about $130 to 180 billion at the low point in 2029. The range depends on how its extra burn falls across the years and whether all of Nvidia's $30 billion has been paid. That is not a crisis; it is a funding plan. But it means OpenAI's bills are paid by its next investors, not its customers, until about 2030.

If growth comes in at half the plan, OpenAI needs about $250 billion of new money by 2030. If revenue stalls at today's level, about $370 billion. A slowdown lets it trim spending at first, so the gap opens later rather than sooner, but about 80% of its planned computing is already signed and the later bills still come due.

Anthropic largely passes. On its plan it builds cash, and if revenue stalls at $65 billion a year it roughly breaks even through 2028. In the harshest case, stalling while paying its contracts at about $100 billion a year, it ends 2028 with about $35 billion of spendable cash and would need its IPO money during 2029. The IPO, expected to raise about $100 billion, covers that.

So the system has one load-bearing wall, not two: OpenAI's ability to keep raising money, roughly $100 billion every year or two until 2030. If that stops, the losses land first on Oracle and the lenders behind its Stargate sites, then on SoftBank, CoreWeave and Cerebras, and from 2028 on Nvidia through its guarantees. Microsoft and Amazon lose future revenue and paper gains, not their footing.