Humanity can't afford not to spend $800B on AI, Anthropic insists

Anthropic filed its paperwork to sell shares to the public for the first time (an IPO) on Monday morning. Its IPO filing and prospectus read like a real estate agent trying very hard to sell a house of cards as a good investment.

The possible benefits are huge, but investors should know about a few problems. The foundation is cracked, the house is in a flood zone, and you can only really afford it if you get a promotion. Still, it might work out in the end, and the view is great.

The sections about risks in IPO filings often sound like warnings of the end of the world. Anthropic, like any company planning to sell shares to the public, must explain the risks it faces. But the company's success still depends on a big gamble: that its income will grow fast enough to cover its rapidly rising computing costs. Over roughly the next decade, its payment commitments now add up to more than half a trillion dollars.

Burn, baby, burn

At first glance, the numbers in Anthropic's S-1 filing do not look too bad for the company behind Claude. In the first half of 2026, its revenue was nearly twice its computing and infrastructure costs. Its operating loss was $693 million, much lower than the $2.14 billion loss in the first half of last year. The company is still losing money, but things seem to be moving in the right direction.

That would be good news if Anthropic's computing costs were expected to stop growing over the next few years. But they are not.

"Our ability to develop, train, improve, and deploy frontier models depends on our access to enormous and increasing quantities of specialized compute," the prospectus says.

Anthropic needs more and more computing power to keep developing its AI models. The company has not given one final total for its commitments to large cloud providers, newer cloud companies, and equipment suppliers. But, depending on how the numbers are counted, its payments over the next few years could exceed $800 billion.

According to the S-1, after December 31, 2025, Anthropic signed new computing and infrastructure deals or changed existing ones, creating about $417 billion in future minimum payment commitments. Since June 30, 2026, it has signed or changed more deals that add about $245 billion in future minimum payments. Some of these payments depend on certain conditions in the contracts being met.

But adding up all the contracts listed gives an even larger total.

• Microsoft: "at least $44.9B" — 2026-2033

• Google: "at least $114.5B" — 2026-2033

• TPUs from Google and other providers: "approximately $125B to $150B" — 2028-2031

• Amazon: "at least $110B" — 2026-2036

• SpaceXAI: "up to $120.5B" — 2026-2029

• AMD: "to exceed" $20B — initial tranche starting in 2027

• Nscale: "up to approximately $44.6B" — through 2033

• Lambda: "approximately $41.3B" — through 2034

• Broadcom: "approximately $161.2B" — starting in 2026

These contracts add up to between $782 billion and $807 billion in commitments, although some deals are more binding than others. Anthropic's S-1 says its agreements with Microsoft, Google, Amazon, and Broadcom cannot be canceled. The contracts with Microsoft, Google, and Amazon also require Anthropic to pay even if it does not use the computing power.

Anthropic can leave its contracts with Nscale and Lambda, which run through 2033 and 2034, only if those companies fail to meet their commitments on time or break the contracts.

Anthropic has more freedom with its widely reported deal with Elon Musk's SpaceXAI to rent computing power from its Colossus supercomputer in Memphis. The deal is worth about $120 billion through 2029. Starting in early 2027, either side can cancel it with 90 days' notice. Anthropic's deal with AMD is expected to cost more than $20 billion for the first batch of computing power, if Anthropic obtains the full amount. However, it can meet this commitment by renting computing power from companies such as Microsoft.

It is still unclear when Anthropic must make all these payments, as several relevant fields in the filing are blank. But the company will need to grow at an extraordinary rate to keep up with its spending, let alone start making a profit.

Compute = revenues, except if it doesn't

All this spending depends on the idea that more computing power will lead to better, more attractive AI models. These models are expected to attract more customers and support products with higher profit margins, meaning the company keeps more money from each sale. That is why Anthropic is telling investors it will keep signing new computing contracts and changing existing ones. At this rate, we would not be surprised if the AI developer's commitments through 2040 passed $1 trillion before the end of this year.

But whether Anthropic can actually meet all these commitments is another question. "Our ability to meet the substantial and growing financial obligations of these requirements to secure compute depends on our ability to generate significant revenue growth," the filing says.

If Anthropic's income does not grow fast enough to cover its computing commitments, it could fail to make its payments. The company warns that this could start a chain of problems that brings down the whole business.

Anthropic has some backup options if it cannot pay its bills using its own money. For example, its S-1 says that "we have the ability to issue, if needed, up to $42.0 billion in aggregate principal amount of convertible promissory notes to Broadcom." These are promises to repay borrowed money that may be converted into company shares. However, this option only covers Anthropic's commitments to Broadcom.

The company could also raise money by selling shares or borrowing to pay its other debts. But if it fails to meet its revenue targets, Anthropic warns that these choices could reduce existing shareholders' ownership or force the company to accept loans with unfavorable terms.

My friend, my enemy, my patron

There is another problem: some of Anthropic's biggest computing suppliers are also among its biggest customers and strongest competitors. These include major cloud companies such as Google, Amazon, and Microsoft. They are much larger than Anthropic and already make profits.

This creates a special challenge for Anthropic. Renting computing power from cloud providers means it does not have to pay for the equipment upfront. But if Google, Amazon, or Microsoft cannot provide enough computing power because they are using it to train or run their own AI models or models for other customers, Anthropic may have to find other options. One possibility would be to build more of its own computing equipment and facilities.

"We may in the future finance a greater portion of our compute on an up-front basis, including through capital leases, prepayments, debt-financed purchases of equipment, or similar arrangements," the company explained.

Anthropic is already doing something like this through its leasing deal with Broadcom. The deal appears to be linked to an arrangement announced earlier this year in which Google and Broadcom worked with Anthropic to provide computing power using Google's tensor processing units (TPUs), special chips designed for AI tasks. Meta was also reported to be discussing plans to install Google's TPUs in its own data centers, so this kind of arrangement would not be unusual.

There are also outside risks. For example, a conflict in the South China Sea could break supply chains that provide most of the world's AI accelerator chips.

The only way to justify spending this much and taking these risks is to believe the possible benefits are enormous — perhaps leading to the biggest economic change in human history.

As Anthropic puts it: "We believe this abundance of AI will rapidly transform every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet." ®