On Monday, October 26, Anthropic filed the paperwork to become a public company. It plans to list on the Nasdaq under the ticker ANT, possibly as soon as Thanksgiving week.
Inside the filing is a number worth stopping on. The company has committed to pay for up to $750 billion of computing power in the coming years.
For scale, the four largest technology companies plan to spend somewhere above $700 billion building data centers this year. We went through those figures in an earlier issue.
So one private company has committed to an amount of computing roughly the size of what the entire industry plans to build in a year. And the companies it has committed that money to are the same companies doing the building.
That arrangement is what people are gesturing at when they use the word "circular." It is worth laying out carefully, because the structure is real and the conclusions people draw from it mostly are not.
What the filing actually says
First the plain numbers, because they are striking on their own.
Revenue for the first six months of 2026 was $16.3 billion. In the same six months a year earlier it was around $1.1 billion. That is roughly thirteen times larger in a year.
Net loss for those six months was $80.7 billion, and this part needs an explanation because the headline is misleading. About $72.4 billion of that loss is a non-cash accounting charge connected to convertible notes the company issued.
Here is how that works. A convertible note is money raised that can later turn into shares. Accounting rules require the company to revalue those notes as its own value rises. When the company is worth more, the notes are worth more, and the increase is recorded as an expense. So a rising valuation mechanically produces a larger reported loss, without a dollar leaving the building.
That is not a trick and it is not hidden. It is just a number that means almost the opposite of what it looks like.
Spending on computing was close to $9 billion in the first half of this year, against $7.3 billion in all of 2025.
Customer concentration: about 20% of revenue comes from two customers.
What a noncancelable commitment is
Now the part that matters for the rest of this.
The filing describes cloud agreements of roughly $111 billion with Google, $110 billion with Amazon, and $31 billion with Microsoft. Much of it is payable whether or not the computing gets used.
That structure has a name in older industries. A take-or-pay contract means you agree to pay for a quantity whether you take delivery of it or not. Pipelines, power plants and mines have used them for decades, and the logic is straightforward. The supplier is being asked to build something enormous and specific, and will not start unless somebody underwrites the revenue. The commitment is what makes the construction financeable.
So these contracts are not unusual in kind. What is unusual is the size, and who is on each end.
The circle
Here is the loop, in four steps.
One. An AI company commits a very large sum to three cloud providers, much of it payable regardless of use.
Two. Those cloud providers are spending enormous amounts building data centers. Alphabet has guided to $195 to $205 billion this year, Amazon to roughly $220 billion, Microsoft to something in the region of $175 billion to $190 billion depending on which source you read.
Three. When investors ask whether that construction is justified, one of the main answers given is backlog. Alphabet reports a cloud backlog of $514 billion. Amazon's cloud unit reports $496 billion. Those are contracted customer commitments, and they are cited as evidence that the demand is real rather than hoped for.
Four. Part of that backlog is commitments from AI companies, which are themselves loss-making and funded by investors, in several cases the same investors who own the cloud providers.
Read that loop quickly and it sounds damning. Read it slowly and it is more interesting than that.
Why this is not automatically a problem
The honest version has to include this part, and most coverage will skip it.
A commitment is a real obligation. Anthropic is on the hook for these payments. If the money is not there, that is a problem for Anthropic, not an accounting illusion. A contract that binds one party and not the other is not circular, it is just a contract.
The revenue is real and it is growing fast. $16.3 billion in six months, from about $1.1 billion a year earlier, is not a rounding error or a projection. The company separately told investors that third-quarter revenue was $16.8 billion, up 45% from the second quarter. Those are customers paying money for a product.
Vertical arrangements are normal. Suppliers invest in customers, customers prepay suppliers, and large buyers get equity in the firms they depend on. This happens constantly in semiconductors, in aerospace, in energy. It becomes a problem when the money going round is the only money, and that is a question about proportion rather than structure.
The thing worth watching is therefore narrower than "is this circular." It is: what share of the cloud backlog comes from companies that are not yet self-funding, and is that share rising?
That number is not published. Which is the actual story.
What this does not tell you
It does not tell you the valuation is right or wrong. Press reports put the listing somewhere between $1.5 trillion and $2 trillion. That figure is not in the filing, it comes from people describing private conversations, and it will move before anything lists.
It does not tell you the commitments will all be paid. "Up to $750 billion" is a ceiling across future years, not a bill arriving next quarter, and large contracts get renegotiated all the time. The filing lists the commitments as a risk, which is the correct place for them.
It does not tell you the loss means what the headline says. Strip out the $72.4 billion accounting charge and you are looking at a very different company. Strip out too much and you are looking at a company spending $9 billion on computing in six months against $16.3 billion of revenue. Both of those are true and neither is the whole picture.
And it does not tell you this is a bubble or is not one. That question cannot be settled by a filing. What a filing can do is give you the specific thing to watch, and in this case it has: whether the companies committing to the backlog can eventually fund those