Anthropic Prospectus Breakdown: $637 Billion Compute Commitment and the IPO Window

In Anthropic's prospectus, the US tech giants may be the biggest winners.

In the risk disclosures, Anthropic mentions a "take-or-pay" model with the giants, meaning an agreed minimum payment amount regardless of actual usage.

The prospectus mentions roughly $637 billion in compute commitments. Among them, Broadcom, SpaceX, Google, and Amazon each reach the hundred-billion-dollar level.

As of July this year, Anthropic's ARR (annual recurring revenue) was only $65 billion, an order of magnitude short. Anthropic is essentially handing over the money of the coming years to the tech giants.

The best valuation window

This is a very good listing window for Anthropic. There are mainly three aspects.

First, Anthropic is in a state of rapid sprinting, and its current growth rate is what can most impress investors.

According to the prospectus, from 2023 to 2025, Anthropic's revenue grew from $41 million to $4.585 billion. In the first half of the year, Anthropic's revenue was $16.286 billion, up 13 times year over year. It seems to have grown rapidly already, yet the third quarter was still explosive; based on information mentioned in the prospectus, third-quarter revenue is estimated at about $16.5 billion, already surpassing first-half revenue, still at a 10x growth rate year over year.

The details shown by ARR also cross-corroborate this conclusion. At the start of 2025, Anthropic's ARR was only $1 billion; by April this year, it had surpassed $30 billion. According to media reports, by the end of July it had reached $65 billion—a growth rate that may be rare even in human history.

Second, Anthropic remains the most user-recognized company in the AI world. This can be seen from user data. Anthropic counts customers on a TTM (trailing twelve months revenue) basis, dividing customers into under $100,000, above $100,000, above $1 million, and above $10 million.

As of September 30, customers above $100,000 reached 11,900, an increase of 10,400 over three quarters. Those contributing more than $10 million had reached 250, while those below $100,000 reached as many as 40 million.

How imperious is Anthropic toward its customers? The prospectus mentions that part of its revenue comes from "customers who violated usage policies or commercial terms." That is, customers whose accounts were banned; even if these customers topped up money, they may not get refunds. Many mainland users may already have felt this, and may even be full of complaints. Even so, many customers still cannot do without Anthropic.

The core reason is that Anthropic's technical capability remains leading. This is also an important "selling point" for Anthropic's current listing.

Finally, Anthropic at this stage proves that technological leadership can tell a coherent business story. In the first half of the year, Anthropic's net loss was $80.748 billion, but this was mainly determined by financing; losses from convertible bonds, derivative liabilities, and fair value changes accounted for $72.388 billion, while the actual operating loss was only $693 million.

If measured by operating profit, Anthropic's second-quarter figure was $974 million, achieving positive operating profit for the first time in a single quarter. In the first half, net cash flow from operating activities was $3.929 billion, achieving positive cash flow, meaning the large-model model has strong profitability.

Listing at this moment, Anthropic can obtain a good valuation. Users recognize it and have a strong willingness to pay, and its market bargaining power is strong. It is in a stage of rapid sprinting, and its high growth rate is very eye-catching. At the same time, it has shown the capital market the sustainability of its business model—it is only temporarily in an investment period, and the business model holds.

If it could catch the window when SpaceX listed, it might be even better. The latter, at its IPO in June this year, with the pie in the sky drawn by AI, saw its valuation once surge to $1.77 trillion.

Now, according to Reuters, some investors who attended Anthropic's investor day have lowered their IPO valuation expectations, from the previous $2 trillion to around $1.5 trillion. The core reason is that the company's quarter-over-quarter revenue growth has slowed. According to previous estimates, Anthropic's quarter-over-quarter growth in the third quarter has already fallen to about 43%.

Now may already be the best timing; otherwise, it would mean waiting longer, more uncertainty, and possibly even worse.

The AI world changes very rapidly. Before Anthropic, OpenAI was the leader in this market. The latter created the ChatGPT moment and maintained leading model capabilities for a long time. In 2021, seven former OpenAI members left to start businesses and founded Anthropic. In just five years, it achieved an overtaking of OpenAI.

And it was an all-around overtaking: in May, Anthropic's valuation reached $965 billion, surpassing OpenAI. In July, ARR surpassed OpenAI.

But Anthropic's lead is also narrowing. Data from October shows that the performance gap between top Chinese and US AI models has narrowed to about 3%, a new low. Chinese large models are continuously catching up.

On the AI model evaluation platform LiveBench, DeepSeek V4.1 Flash ranks sixth globally. Its overall performance is already comparable to the leading models from Anthropic and OpenAI. Moonshot AI's Kimi K3.1 Pro has performance approaching that of Opus 5.5.

Chinese models cost less and offer better value for money, and on both the technology capability and commercialization tracks, they could deal a blow to Anthropic. More critically, Chinese models are more open and will not be subject to global export controls like Fable 5.

Perhaps at some point in the near future, Anthropic will no longer lead, and its commercial monetization will also face a huge impact. In addition, under rapid growth, the base will keep getting larger, and a slowdown in growth rate is itself unavoidable; the later the valuation, the more likely it may be inferior to the present.

To avoid this situation as much as possible, Anthropic has taken countermeasures.

Anthropic has cut prices to a certain extent, especially for small models. In October, it released Claude Haiku 5.5, its smallest and most affordable model, with input and output unit prices significantly lowered compared with the previous generation: the standard input price dropped to $0.10 per million tokens and the output price to $0.50 per million tokens, a unit price reduction of about 90%.

Flagship models have also seen price cuts, though only to a small extent. For example, in September, Anthropic released Fable 5.1, priced at $10/$50/$0.25 (input/output/cache read, per million tokens, same below). It is the same as Fable 5 in June, but Fable 5.1 reduces the cost of typical workloads by about 25% compared with Fable 5, with savings of up to 45% on complex coding and high-intensity agentic tasks, which also counts as lowering the cost of using the model.

Before submitting its prospectus, Anthropic released Claude Opus 5.5, priced at $4/$20/$0.20; after combining model efficiency improvements, the cost per task is about 40% lower than Opus 5.

But such prices are vastly different from those of domestic Chinese large models. For example, the newly released DeepSeek Flash V4.1 1017 costs $0.15 per million tokens for input, $0.60 for output, and $0.003 for cache read. Compared with Anthropic, it is almost dirt cheap. Kimi K3.1 Pro costs $2 for input, $10 for output, and $0.20 for cache read, with input and output prices only half of Opus 5.5.

Anthropic's price cuts are only half-hidden, and its prices are still too high compared with Chinese large models. This is determined by cost and business logic. Even in Anthropic's codebase, in September alone, more than 90% of new contributions to Anthropic's codebase were written by Claude. Costs can be pushed down, but its prices most likely cannot reach the extreme levels of Chinese models.

Therefore, for a long time to come, Anthropic will have no price advantage.

So, if it wants to maintain a global leading advantage, there is only one path—technological leadership.

A dazzling array of capital maneuvers

The prerequisite for technological leadership is computing power security. Anthropic has become a major financier of giants such as Broadcom, SpaceX, Google, Amazon, and Microsoft.

SpaceX's third-quarter report shows that its AI business revenue reached 7.56 billion yuan, up 830% year on year, surpassing launch and communications to become its largest business. SpaceX did not mention too many customers; Anthropic is the only one mentioned and also the largest external customer.

In the prospectus, Anthropic mainly mentions seven companies. Among them, Broadcom is the largest source of procurement; the two sides signed a five-year master equipment finance lease, with rent of about 161.2 billion, paid quarterly.

SpaceX's total procurement amount is at most 120.5 billion, with cooperation through 2029, and orders before 2027 are irrevocable. After the deadline, either party may cancel with 90 days' notice. This counts as the cooperation with the most flexible exit.

Google's computing power services amount to as much as 114.5 billion; it is a typical case of no matter the usage, shortfalls to be made up, and irrevocable except for major breach by Google. Not only that, Anthropic also committed to additional TPU purchases of 125-150 billion in 2028–2031.

Amazon's computing power services are 110 billion; like Google, it is also no matter the usage, shortfalls to be made up, and irrevocable except for major breach. Microsoft's is 44.9 billion, no matter the usage, partly revocable.

With just these giants, Anthropic has committed to procurement of more than 500 billion at most. Even at an ARR of 100 billion, Anthropic has basically handed over its revenue for the next five years to the giants; this can basically be said to be working for the giants.

And Anthropic does not have enough funds. As of the end of the first half, Anthropic had 112.4 billion in cash and equivalents and 19.3 billion in short-term investments, far from enough to cover the huge procurement orders.

Therefore, Anthropic can only solve this problem through financing.

The most typical is the NPA (note purchase agreement). Briefly, it means that through financing, relevant funds are drawn conditionally. For example, the NPA with Google stipulates a TTM revenue target. The cooperation with Broadcom sets requirements for revenue growth. Amazon's 15 billion NPA is unlocked based on delivered computing power: only as much computing power as Amazon AWS delivers can Anthropic draw down.

Second is convertible bonds, mainly the 6.2 billion associated with Amazon, with the conversion price taken as "the lower of the fixed price and the IPO price."

To avoid control also being carved up by the giants, Anthropic adopted five equity models. What is publicly offered is Class A, with economic rights. The founder team holds Class F, with no economic rights but 50.1% voting power on ordinary voting matters. The conversion channel is Class N, employee incentives are Class E, and Class T has the right to elect and remove the board of directors.

It can be seen that those with decision-making power over the company are the seven founders, while external tech giants have not been listed as shareholders and will all have shareholding caps in the future.

Through such methods as financing, conditional unlocking, and dual-class share structures, Anthropic maintains its computing power supply as much as possible, which in fact also keeps the entire AI industry mired in doubts about "self-entertainment."

Setting aside these clouds at the capital level, Anthropic's real value still lies in customer count and ARR metrics.

Looking only at these metrics, Anthropic is a phenomenal company in a phase of rapid expansion. What is special is that it is in the AI industry, an industry that may determine the course of humanity, and Anthropic is the king of this industry. But this king's leading advantage is not eternal. At this stage, it may instead be the best time to go public.

This article comes from the WeChat public account "Shudushe," author: Rousseau, published by 36Kr with authorization.