Anthropic’s amended filing lists $17 billion in quarterly revenue and $895 billion in commitments

Read past the revenue number. Anthropic's amended filing is a confession dressed as a growth story.

The growth is real: $33 billion in nine months, up roughly 1,200%. But the filing also lists about $895 billion in purchase commitments and leases, nearly all for computing and infrastructure. Third-quarter revenue of $17 billion annualizes to around $68 billion. The company has promised to spend more than thirteen times that on chips, data centers and power. Compute costs rose 329% in nine months, and growth slowed from 144% sequentially in the second quarter to 47% in the third. That is still fast, but the promises are made as if it will never slow further.

The $104 billion net loss deserves honesty. The filing attributes it mainly to the changing fair value of convertible notes and derivatives, not to cash burned. That is an accounting artifact of soaring valuations. It is not proof of insolvency. But it is proof of something else: a company whose paper value is swinging by tens of billions on investor sentiment, and which has obligated itself to nearly a trillion dollars on the assumption that sentiment holds.

The usual answer is that this is private money, so who cares? Customers choose to pay. Investors choose to invest. If it fails, they eat the loss.

That would be convincing if this were a software start-up. A trillion-dollar buildout is a public project with private ownership. It claims land, water and enormous quantities of electricity from grids that ordinary households also depend on. When utilities build generation and transmission for a few giant customers, the costs of that capacity can end up in everyone's rates. The debt and lease structure ties lenders, chipmakers, landlords and eventually pension funds to the same bet. If the bet sours, the same executives who sold this as civilizational infrastructure will argue that it is too important to let fail. We have seen that script.

Now the 90 pages of risk disclosures, which cover political risk, infrastructure constraints and the tension between safety and commercial goals. They are the most candid account of the company's own dilemma we are likely to get, and we have it because securities law compelled it. That is the point. Anthropic's founders say they want to explain and reduce the risks of AI while working with governments and peers. Fine. But a firm that tells investors its safety mission competes with its commercial incentives has told the public something essential: voluntary restraint will lose to the quarterly race.

Anthropic has long made a collective-action argument itself. Every lab fears being outrun, so everyone sprints. The Left should take that argument more seriously than the company's own politics do. If the people building the systems say competition bends their judgment, then the answer is binding public rules, not better branding. Those rules would cover energy and grid costs, transparency about model risks, incident reporting and worker protections. Instead, Washington's instinct has been to shield the industry, including repeated attempts to strip states of the power to regulate AI at all.

The Left has its own failing here. Too many Democrats treat AI as either a culture-war nuisance or an economic trophy for their districts. Governors court data centers with tax abatements and secrecy agreements, while ratepayers learn the terms later. If we want a different outcome, we should demand that every subsidy, power contract and zoning exemption come with public disclosure, community benefit and a guarantee that residential customers do not underwrite corporate capacity.

The filing shows a company that sees the stakes. It lists them in plain language across ninety pages. The question is whether the rest of us will let a handful of boards and investors decide how those risks are distributed, or whether democratic institutions will insist on a say before the bill arrives.