Anthropic’s 800-billion euro time bomb
The company behind Claude is preparing to make the largest stock market debut in history. It’s also opened its books to the public for the first time. Although the numbers are impressive, the outlook’s still unsettling.
Anthropic is preparing for its initial public offering. The leading AI company has filed the required S-1 form with the US Securities and Exchange Commission (SEC). The report, comprising over 300 pages, discloses detailed information about Anthropic’s finances for the first time ever.
By contrast, the exact time of the initial public offering (IPO), the planned number of shares and their price haven’t yet been defined. Analysts expect Anthropic to aim for a valuation of 1.5 to 2 trillion US dollars and make over 100 billion. This would outdo SpaceX as the largest initial public offering in history.
At first glance, the books show strong revenue growth and smaller losses than previously feared. But upon closer inspection, they also reveal massive financial commitments and investment plans. In other words, Anthropic has put itself into a high-risk position. Over the next few years, demand for Claude will have to keep on skyrocketing to prevent the company from imploding.
Anthropic in a stronger position than OpenAI
During the first half of 2026, Anthropic generated 16.3 billion dollars in revenue. That’s 14 times more than in the same period last year. Operating costs totalled 17 billion. This resulted in an operating loss of 0.7 billion (-4 per cent operating margin). During the first half of 2025, the loss was still 2.1 billion (-184 per cent operating margin). For the second quarter of 2026, the company even reported its first operating profit of 1 billion (8 per cent operating margin).
The numbers are better than expected. It seems, Anthropic was able to scale its business more sustainably than its largest competitor, OpenAI. According to leaks, the latter spent a staggering 34 billion with a revenue of 13 billion, resulting in an operating loss of 21 billion (-162 per cent operating margin). Although the comparison is offset in time and covers different periods, it still illustrates how the relationship between costs and revenue differs.
It seems Anthropic is more efficient at developing and operating its models. The total costs for the items «Computing Power and Infrastructure» (8.9 billion) and «Personnel»(3.1 billion) amounted to «only» 12 billion in the first half of 2026. For the same items, OpenAI incurred a total of 27.3 billion in 2025, despite lower revenue. These are broken down into 19.8 billion for «Research and Development» and 7.5 billion for «Operating Costs.»
Glossed-over books, but no skeletons in the closet
It’s impressive that Anthropic managed to increase its efficiency so significantly. Almost a little too impressive. The spikes in the first and second quarters of 2026, in particular, need to be taken with a pinch of salt. In the fourth quarter of 2025, the company generated USD 0.69 in revenue for every dollar spent on computing costs. In the first quarter of 2026, it was at USD 1.40, and as high as USD 2.07 in the second quarter. Analyst Ed Zitron assumes these figures can be explained by a discount granted by SpaceX, from which Anthropic purchases computing power. This means it’s likely that Anthropic’s efficiency will look less impressive in the third quarter.
The figures published to date are in line with generally accepted accounting principles (GAAP). For some of its charts, however, Anthropic has been window-dressing its figures: «Adjusted Operating Profit» is stated to be 2.2 billion in the second quarter of 2026, and the «Adjusted Operating Margin» 19 per cent. To generate these figures, the accounting department has deducted the equity-based compensation for employees, among other things. In other words, the part of the salaries Anthropic pays out in stock options instead of cash.
Other major tech companies, such as Meta, have been using the same accounting trick for a while now. But it’s all smoke and mirrors. Although stock options don’t have a direct impact on cash flow, they do once employees make use of them. When that happens, the company must buy back the same number of shares on the stock market using cash. Otherwise, the shares of remaining investors are diluted, which would cause the stock price to fall. «Adjusted Operating Profit» is simply a short-term way of making the numbers look better and a trick that eventually catches up with you.
An even larger negative item, which is also not considered in the ordinary operating income, consists of revaluations of convertible notes. In the past, Anthropic has been using these to raise funds from Amazon, Google and Broadcom. The notes can be converted into shares upon maturity, which is now happening on a large scale.
Since the company’s valuation has skyrocketed in the meantime, Anthropic is incurring book losses. In the first half of 2026, these amounted to a whopping 72.3 billion dollars, resulting in a net loss of 80.7 billion. However, these book losses have no impact on cash flow, not now or in the future. Seeing that the majority of the convertible notes will be converted before the initial public offering, they’re unlikely to dilute the stock price down the line. This means net loss is the wrong metric for assessing operating profit.
Over 800 billion euros (!) of financial commitments
The true time bomb Anthropic needs defuse isn’t hidden in its past figures, but in the sections on contractual commitments and future investments in infrastructure.
To date, Anthropic has hardly any servers of its own. Instead, the company trains and runs its AI models on Amazon, Google, Microsoft and SpaceX. It has long-term contracts with these providers, regardless of whether the computing power is actually used in the end. This is a lucrative business for the hyperscalers, as they can charge Anthropic a fixed margin on their costs.
It’s quite telling that the total amount of these «Computing Commitments» aren’t disclosed in the S-1 filing. The relevant sections have been redacted in the public version. However, details revealed about individual contracts are suggesting that the figures are staggering. Fortune and Reuters estimate the figure at over 500 billion, while the New York Times has put it at over 800 billion over the next ten years. Most of it non-refundable.
Our increasing efforts to build out our compute and infrastructure require substantial capital expenditures and are changing the cash flow profile of our business.
To ensure the AI supplier is no longer dependent on Microsoft and others in the long term, it also plans to build its own servers, as stated in its prospectus: «Our increased efforts to expand our computing capacity […] require substantial capital expenditures and are changing the cash flow profile of our business.» This entails «significant upfront payments well in advance of the related compute becoming operational.»
In other words, Anthropic will be needing vast amounts of cash over the next few years. It has no choice but to purchase hundreds of billions of US dollars’ worth of computing power from competitors, while simultaneously building its own servers. And they won’t pay for themselves until later.
IPO at a strategically perfect turning point
All in all, the planned expenditures are so gigantic that the financial figures to date are virtually insignificant. The future will look completely different. The IPO is perfectly timed to coincide with a turning point. Namely, just as Anthropic is still reporting rapid growth and even an operating profit compared to the previous year, but just before the lion’s share of its massive commitments and investments comes due.
In the first half of 2026, spending on computing power totalled 8.9 billion US dollars – peanuts compared to the promised 500 billion (or more) over the next ten years. To cover these costs from its ongoing operations, demand for Anthropic’s AI products would have to continue to skyrocket. However, leaked figures show that revenue growth is already slowing down. In the first two quarters of 2026, demand was still 132 and 144 per cent, respectively, but in the third quarter it had fallen to just 45 per cent. This reflects the end of the era of tokenmaxxing indicated by the red bars in the chart below (linked article in German).
In the best-case scenario, things will still pan out in the long run. For that to happen, growth would have to pick up again, or at least not slow down. At the same time, computing costs per token would have to decrease, or customers would have to pay higher prices for plans. That way, Anthropic could make use of the hundreds of billions worth of computing power it’s already purchased and turn it into profit thanks to a positive margin.
Opinions are divided on how realistic this scenario is. In the short term, Anthropic will definitely be needing a lot of money from capital markets. According to the S-1, the company currently has surprisingly large cash reserves of 112.4 billion. The initial public offering is likely to roughly double this figure. But even that amount will run out fast given the computing cost commitments and the cost of constructing their own servers.
To raise more cash, Anthropic could issue additional shares at a later date. However, that would dilute the stock price every time and would only work as long as Wall Street believes in the company’s success. Another alternative would be to issue bonds. But the market now asks for high interest rates from AI startups for issuing bonds, as it considers the default risk to be increasingly high. For example, Neocloud supplier CoreWeave already pays 15 per cent per year for new capital. If Anthropic were to finance itself this way, it would incur high operating costs.
The bottom Jenga block
The weight of Anthropic’s growth rests not only on the company itself, but on the entire AI industry. The hundreds of billions in financial commitments have been factored into the recipients’ growth forecasts. So if Anthropic fails to pay, Microsoft, Google and Amazon will lose that revenue. It would also mean a reduced need for new servers and GPUs from Nvidia.
At the same time, Anthropic’s success is viewed as a litmus test for the business models of many AI suppliers. If the market leader fails to achieve profitability, it fundamentally calls into question whether the costly development of Frontier models is financially worthwhile – and whether it justifies the trillions in investments made to date and planned for the future. Anthropic is the allegoric bottom block of a giant Jenga tower.
Source: Shutterstock
Nevertheless, Wall Street’s breathing a sigh of relief for now. Many had expected worse results. Concerns about an AI bubble have been negatively impacting tech companies’ stock prices in recent weeks. But after the release of Anthropic’s S-1, they started to pick up again. Looks like the party’s not over yet.
Update 4 November 2026: meanwhile, Anthropic’s released a new version of the S-1 filing. It contains its third-quarter figures and the official total of financial commitments are no longer redacted:
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.
Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.
Show all
