On Monday, Anthropic filed paperwork for what could end up being the largest initial public offering in history. The registration document, no less than 335 pages, offers the public a rare opportunity to learn about the inner workings of one of the leading providers of cutting-edge AI models. The text highlights both the immense potential of this technology, still in its infancy, and the almost incalculable risks involved in taking AI ever further.
On a financial level, the document reveals exponential growth, heavy losses and a clear path to profitability. Anthropic’s revenue went from $41 million in 2023 to $16.3 billion in the first half of 2026, with an increase of 1,300% between the first six months of 2025 and 2026 alone. And, although the document includes net losses of $80.7 billion in the first half of 2026, these were mainly due to changes in fair value of convertible debt securities, that is, to factors unrelated to operational activity. In fact, the company recorded an operating profit of $974 million in the second quarter of 2026, after growing its revenue at a rate much higher than its operating expenses – especially computing and infrastructure – during the last 18 months.
Perhaps the most significant revelation in Anthropic’s IPO filing is that the company has accumulated up to $800 billion in computing and infrastructure spending commitments. As our chart, which includes only the largest deals, shows, Anthropic has signed more than $100 billion in deals with partners such as Broadcom, Google, SpaceX.AI, and Amazon, most of which cannot be canceled. The magnitude of these commitments, coupled with the fact that a large portion of Anthropic’s revenue is consumer-dependent and therefore variable, poses a significant risk: the company could quickly find itself in difficulty if the number of customers and usage of its services does not continue to grow as expected.
“We must expand our customer base and rapidly grow revenue to generate the cash flows necessary to meet these commitments,” Anthropic wrote in its S-1 filing. “However, our revenue depends primarily on consumption and much of it is not supported by long-term contractual commitments, while our computing and infrastructure commitments typically span several years. Therefore, we may assume these obligations based on expectations of income that may not materialize.