Anthropic loses $80.7 billion in six months, but the key to its IPO prospectus is $600 billion in committed compute

By Benzinga (with reporting from the Wall Street Journal) · October 20, 2026.

Anthropic posted a net loss of $80.7 billion in the first half of 2026, compared with $13 billion in the same period of 2025. The figures come from the draft prospectus for its initial public offering (IPO), which The Wall Street Journal obtained, according to Benzinga. Yahoo Finance also reported on the same document.

The figure is alarming, but most of it is not cash burned in operations. About $72 billion corresponds to accounting charges for changes in the value of convertible notes and other financial instruments issued to raise capital. These are fair-value adjustments (revaluations of liabilities): the more the company is worth, the more, on paper, it owes to those who financed it earlier. For full-year 2025, the net loss was $42 billion, and more than $34 billion of that was also adjustments of this kind, according to the prospectus.

The operating numbers are what show the scale of the business. Revenue for the half was about $16.3 billion, compared with $1.1 billion a year earlier. That is roughly 15 times more. Spending on compute and infrastructure rose to almost $9 billion, from $2.3 billion in the first half of 2025. In those six months alone it exceeded the $7.3 billion spent in all of 2025, a year in which the company brought in almost $4.6 billion.

The prospectus adds future commitments. Anthropic has scheduled payments of $111.1 billion to Google, $110 billion to Amazon Web Services (AWS) and $31.4 billion to Microsoft, even if it does not use all the contracted capacity. In addition, it could pay up to $86 billion to rent equipment with Nvidia chips from Nscale and Lambda, under contracts running through 2033 and 2034. Including Google's chip-rental deal with Broadcom, the total could exceed $600 billion, according to the WSJ. The company is negotiating other deals that would add another $150 billion.

The document acknowledges several kinds of risk. The company admits that access to compute is uncertain: it has requested more capacity than some major cloud partners were willing to provide, because they increasingly see it as a competitor. It also cites “political resistance” to data centers, which according to the text has become “more organized.” And it warns of safety risks: models exhibiting self-preservation behavior to avoid being shut down, or responses resembling blackmail. A catastrophic incident could cause damage that insurance would not cover or would not fully cover. Separately, the independent evaluator METR maintains that Anthropic's risk assessment does not provide sufficient evidence to confidently rule out serious misalignment.

The document also includes corporate governance details. Dario Amodei received about $18 million in 2025 and Daniela Amodei about $16.5 million, mostly in stock and options. The co-founders have pledged to give 80% of their personal stakes to philanthropy, and the company plans to allocate 5% of its shares to a public benefit fund. The IPO could value the company at up to $2 trillion, according to the published reports.

Our take: the $80.7 billion loss headline is the least informative part of the prospectus. That number rises and falls with the valuation and says little about the business. What matters is something else: revenue that has grown 15-fold and compute commitments of roughly 37 times half-year revenue. A bet of that size only pays off if demand keeps growing at the same pace for almost a decade.

That connects with a thesis we have long argued: AI's bottleneck is available energy and compute, not just model quality. The prospectus confirms it from the inside. Anthropic is paying upfront, and at almost any price, for capacity it is not guaranteed. There is also a competitive paradox: its cloud providers are at the same time its rivals and, therefore, its weakest point.

A caveat when comparing: revenue figures across AI labs are not always calculated using the same accounting criteria, so the numbers in this prospectus should be read with caution until there is a common standard.

On safety, the prospectus and METR's review point in the same direction. A company admitting in a regulatory filing that its models could exhibit self-preservation is not alarmism: it is the risk disclosure the market requires. And it is a good sign that an independent third party audits its tests, but it also shows that safety verification still lags behind capability.

In the short term, the risk is financial and infrastructural: enormous compute debt, local opposition to data centers and dependence on rivals. In the long term, if that capacity translates into models able to accelerate biology and medicine, the investment may be the price of entry to the abundance we anticipate. One example cited by Benzinga: Claude designed hundreds of thousands of candidate proteins against scorpion toxins in 64 hours. But that is a capability still to be proven with clinical results, not an established fact. The prospectus gives an honest portrait of how expensive and fragile the road is.