The money going in circles

On May 5, The Information reported a number that everyone repeated and almost no one interrogated: Anthropic has committed to spend roughly $200 billion with Google Cloud over five years for TPU capacity and services. The headline was the size. The more interesting number is the one the headline buries — Anthropic was running at around a $30 billion annualized revenue rate in April, which means the year-one compute bill is larger than the company's entire current revenue. That is not a contradiction the market is pricing. It is a bet that the revenue catches the compute, and catches it fast.

Strip away the size of the figures and the AI build-out is a simple structure repeated at scale: the companies selling compute are financing the companies buying it, and the companies buying it are committing to revenue the sellers can show their own investors. The money moves in a circle. Most of the time, a circle is just an efficient way to allocate capital under uncertainty. The question worth asking — the only question, really — is what the circle looks like if the demand thesis is wrong.

How the money actually moves

Start with the cleanest version. Google agrees to invest in Anthropic — at least $10 billion, reportedly scaling toward $40 billion if targets are hit. Anthropic agrees to spend roughly $200 billion buying Google Cloud and Google's TPUs, the custom chips Google builds with Broadcom. Google books a customer commitment large enough to move its disclosed cloud backlog; one read of the numbers has Anthropic accounting for more than 40% of the backlog Google showed investors. Anthropic books the compute it needs to train the next Claude. Both sides get to point at the other as validation.

This is vendor financing, and vendor financing is not new or inherently sinister. A supplier extends terms — or, here, equity — to a customer so the customer can buy more of what the supplier sells. It lubricates a market that would otherwise be gated by the customer's balance sheet. The mechanism does real work: Anthropic cannot fund five gigawatts of TPUs out of $30 billion of revenue, so the capital has to come from somewhere, and the most motivated lender is the company that wants the compute bought.

The complication is that the same structure is running everywhere at once, and the nodes overlap. Nvidia has been in talks to take an equity stake in OpenAI reportedly as large as $100 billion — a deal Nvidia itself disclosed in a filing might not come to fruition, and which the Wall Street Journal described early this year as stalled. OpenAI, in turn, has assembled infrastructure commitments that, tallied across vendors, run past a trillion dollars over the coming decade: Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon, CoreWeave. Oracle is raising tens of billions in debt to build the data centers OpenAI has promised to fill. Nvidia sells the chips that go into those data centers. The buyer of the chips is partly funded by the maker of the chips. Bloomberg has mapped the loop in detail, and the map is the story.

Vendor financing is fine until the vendor, the customer and the lender are the same three companies. Then it isn't financing. It's a position they all share and none of them can exit alone.

Concentration, the risk that arrives all at once

Concentration is the risk that compounds quietly and arrives all at once. The four largest US hyperscalers — Amazon, Alphabet, Meta and Microsoft — are guiding toward something close to $700 billion of capital expenditure in 2026, roughly double 2025. A large share of that is power and buildings rather than silicon, which matters, because it is the part you cannot resell or repurpose if the demand does not show. But the silicon and the cloud commitments funnel into a startlingly short list of counterparties. One ecosystem's revenue is another's capex is a third's equity stake. The fortunes are tied together by design.

That is the part the dot-com comparison gets right, and the part it gets wrong. It gets right that vendor financing magnifies losses: when Lucent and Nortel lent telecom upstarts the money to buy Lucent and Nortel equipment, the revenue was real until the customers folded, at which point the receivables and the revenue vanished together. It gets wrong the scale of the balance sheets involved. The companies at the center of the AI circle are among the most cash-generative in history; the top names threw off hundreds of billions in operating cash flow last year. They can absorb a great deal before anything breaks. The risk is not that they are fragile. The risk is that they are all exposed to the same single assumption.

The assumption everything rests on

The assumption is that AI demand grows into the compute being built for it — that the inference revenue, the enterprise seats, the agentic workloads arrive on roughly the schedule the capex assumes. Every node in the circle is, in effect, long that assumption. Anthropic is long it at $40 billion a year. Oracle is long it with borrowed money. Nvidia is long it as both supplier and shareholder. The hyperscalers are long it in concrete and transformers. There is no node that is short, no natural hedge inside the system, no participant whose interest is served by the demand disappointing.

This is the herding problem in its purest form: smart people buying the same thing at the same time and calling it conviction. The buyers and the believers are the same people, and when that is true, the mechanism that carried the thing up has no brakes on the way down. If demand softens — not collapses, merely softens, arriving slower or smaller than the contracts assume — the revisions do not stay contained to one name. They travel the circle.

What breaks, and in what order

Think in exposure, not headlines. If the demand thesis disappoints, the damage does not land evenly. It lands first where the leverage is highest and the optionality lowest.

1. The debt-funded builders go first. A company borrowing to construct capacity against a single customer's forward commitment has the least room. If that customer renegotiates or slips, the interest cost stays and the revenue does not. Oracle's bet is structurally the most exposed link, not the largest.
2. The labs feel it next, through the equity round that does not clear. Anthropic and OpenAI are funding multi-year compute out of capital that must be continually raised. The compute commitments are contractual; the funding is a series of rounds priced on the prior round's optimism. A repricing of the thesis reprices the round, and the gap between committed spend and available capital becomes the whole story.
3. The chipmaker is paradoxically the most cushioned and the most watched. Nvidia sells into all of it, so its revenue is the most diversified across the circle — but it is also the name carrying the most index weight, which means a multiple compression there does the most damage to the passive investor who never chose the position.
4. The hyperscalers absorb the most and disclose the least. Their cash flows can eat enormous write-downs, but their capex is now a meaningful share of the market's total earnings growth. The pain shows up not as insolvency but as the multiple coming in across the whole complex at once.

There is a useful contrast outside the American loop. ByteDance is weighing capital expenditure of as much as $70 billion this year, by Bloomberg's reporting — a 25%-plus jump on the prior plan — but it intends to cover much of it from the roughly $50 billion in profit it earned in 2025. That is the old-fashioned way to fund a build-out: with your own cash, against your own revenue. It can still be a bad bet on demand. But it is not circular. ByteDance is not relying on its suppliers to finance its purchases of its suppliers' products, and it is not pledging revenue that becomes someone else's equity story. When the financing is internal, a demand miss is a write-down. When the financing is circular, a demand miss is a chain reaction.

Separating the company from the trade

None of this is a verdict on whether the technology is good or whether the demand is real. Claude may be worth every gigawatt; the inference workloads may arrive on schedule and the contracts may look conservative in hindsight. A trade can be a great technology and a dangerous position at the same time, and the discipline is to hold both thoughts. The companies are extraordinary. The structure they have built to finance themselves is the thing to watch, because it has converted a set of independent bets into a single correlated one.

The number to keep is not the $200 billion or the $700 billion or the trillion-dollar tally that gets repeated on every earnings call. It is the share of all of it that depends on one assumption being right, financed by the same handful of names taking turns being each other's customer, supplier and shareholder. While the story holds, the circle is a machine for turning conviction into capacity. The morning the story wobbles, it is the same machine running in reverse — and the people who would normally step in to buy the dip are already inside, holding the position they sold to each other. That is the exposure. No one is pricing it, because pricing it would mean admitting they are all on the same side of the same trade.