Anthropic's IPO Filing Shows $582B Owed to Five Companies and Unaudited Revenue

Anthropic filed its IPO registration statement last week ahead of a planned November debut, and the amended document carries a number that should stop the process cold: at least $582 billion in non-cancelable spending commitments to five companies. Amazon, AMD, Broadcom, Google and Microsoft are the counterparties. The cash on hand is $96 billion. Matt Scherer, a fellow at the Open Markets Institute, argues in a Tech Policy Press perspective published November 2 that the SEC should refuse to let the company list until it supplies two consecutive quarters of audited financial statements and discloses all material spending obligations.

We think he is right, and the AI industry should be paying closer attention to why.

The headline numbers in the filing look fine. Revenue appears to have grown quickly through 2026. The balance sheet reads healthy on the surface. Then you reach the footnotes, where the $582 billion to five vendors sits alongside roughly $300 billion more in contractual commitments, for a total near $895 billion, most of it stretching into the 2030s. Scherer notes Anthropic is still negotiating further infrastructure and compute contracts and could end the year above a trillion dollars in commitments. These obligations do not appear on the balance sheet. They are, in his phrasing, effectively debts: payments the company is contractually required to make regardless of whether it can afford them.

That is the part the SEC’s own standard seems built for. The agency can delay or block a listing when a registration statement is “on its face incomplete or inaccurate in any material respect.” Anthropic’s prospectus includes unaudited financial results for the period where it claims its fastest revenue growth, buries its largest future obligations in footnotes, and keeps announcing new eleven-digit commitments as the IPO approaches. The third-quarter results suggest revenue growth slowed to a crawl over the summer. A company asking public markets to price it on unaudited numbers from its best quarter, while its worst obligations live in small type, is not a company giving investors an accurate picture.

The margin math does not close

Scherer’s report with Maya Jenkins, “Brace for Impact,” lays out four forces squeezing model-developer margins, and each one applies directly to Anthropic’s filing.

Demand is the first. One study he cites found the share of Americans using AI at work fell from 46% in summer 2025 to 33% in summer 2026. When Anthropic released its Fable 5 frontier model this summer, the commercial reception was tepid. The spending commitments assume businesses will pay for ever-larger and more expensive models. The evidence points the other way.

Competition is the second. Google, Meta, Microsoft, OpenAI, SpaceX and Anthropic all ship models that are largely indistinguishable in quality, so nobody can raise prices without losing customers. Anthropic and OpenAI launched a price war this summer that is still running. Open-weights models from Chinese developers and models that run locally on phones and laptops push prices down further.

Cost is the third, and it is the one that turns the others lethal. A 1 GW data center costs an estimated $31 billion once you add chips, servers, networking, electricity, cooling and construction. Anthropic has committed to 15 GW of capacity and estimates it will eventually need 20 to 25 GW. That is where the half-trillion-plus in commitments comes from, and it is a fixed cost that does not shrink when demand disappoints.

External constraints are the fourth. Backlash to data centers has scuttled some projects and delayed many others. Scherer flags rising costs from the Iran War and supply-chain strain. Every delay cuts the compute Anthropic is counting on to generate the revenue it needs.

The filing asks public markets to price Anthropic on unaudited numbers from its best quarter while its worst obligations sit in the footnotes.

Who ends up holding the bag

The strongest argument in Scherer’s piece is not about Anthropic’s solvency. It is about who absorbs the loss.

As a private company, Anthropic’s shares sit overwhelmingly with insiders, tech giants and venture capitalists. Once it lists, those shares flow into 401(k)s, retirement accounts and index funds. If the margin squeeze bites, the people left holding the stock are workers and retirees who never chose to underwrite a $895 billion compute bet. Scherer also notes the market Anthropic would join is already historically overvalued and concentrated in a handful of interconnected mega-cap tech names, so adding another large, unprofitable giant raises contagion risk rather than diversifying it.

That framing matters for the AI industry beyond one company. Anthropic’s filing is the first time a leading frontier lab has had to put its compute obligations in front of public investors. The document is a stress test of whether the AI buildout’s economics can survive daylight. If the SEC waves it through on unaudited numbers and footnoted debts, every lab that follows will learn the same lesson: disclose the minimum, list anyway, and let the index funds carry the risk.

What this means for builders

The practical read for anyone building on Anthropic’s models is that the compute commitments are the product roadmap. Fifteen gigawatts of contracted capacity, and a stated need for 20 to 25, is a bet that inference demand keeps climbing and that customers keep paying premium prices for frontier capability. The tepid reception to Fable 5 and the drop in workplace AI use are direct evidence against that bet. Builders should expect pricing pressure to continue in their favor in the near term, because a company carrying $582 billion in obligations to five vendors cannot afford to lose customers to a cheaper rival.

The SEC has a narrow window before the November debut. Scherer’s ask is modest: two quarters of audited financials and full disclosure of the spending commitments. If Anthropic cannot clear that bar, the question is not whether the IPO should be delayed. It is why the public should be asked to fund a trillion-dollar compute bill before anyone has audited the books.

The commission’s decision, whenever it comes, will be the first real test of whether AI companies face the same disclosure rules as every other issuer. Watch the amended filing for audited numbers. If they do not appear, the debut should not either.