Anthropic's latest prospectus disclosure shows that its revenue reaches 16.3 billion, and its book loss is as high as 80.7 billion US dollars.

Among the AI giants in Silicon Valley, Anthropic has long been regarded as the most restrained, security-focused, and most similar to a traditional enterprise software company. However, according to the draft IPO prospectus exposed by media including *The Wall Street Journal*, the company presents itself to the capital market as an "asset-heavy player" that relies heavily on computing power infrastructure and exchanges long-term commitments for model competitiveness.

Several sets of core data in this draft can be regarded as the most eye-catching financial snapshot in the history of contemporary technology:

Revenue in the first half of 2026: 16.3 billion US dollars. Net loss in the first half of 2026: 80.7 billion US dollars, of which about 72 billion US dollars is a non-cash accounting provision related to changes in the fair value of convertible notes and other financing instruments, not actual cash outflow in the current period. Computing power and infrastructure expenditure in the first half of the year: nearly 9 billion US dollars, exceeding the 7.3 billion US dollars for the whole of 2025. Cloud computing commitments for the next 7 to 10 years: about 450 billion US dollars, most of which are non-cancellable, and about 80% of the amount needs to be paid "whether used or not". The committed amount for renting NVIDIA chips: up to 86 billion US dollars, with contract parties being AI cloud service providers such as Nscale and Lambda, with terms expiring in 2033 and 2034 respectively.

Looking only at the half-year revenue of 16.3 billion US dollars, it is quite impressive for a young AI company that has not been established for a long time. However, the subsequent book net loss of 80.7 billion US dollars and the rigid 10-year cloud bill still constitute huge financial pressure. It should be noted that the above 450 billion US dollars cloud computing commitment and 86 billion US dollars chip lease add up to about 536 billion US dollars. If the approximately 161.2 billion US dollars equipment lease obligations related to Broadcom are included, the total disclosed infrastructure commitments of Anthropic exceed 600 billion US dollars; if the agreements still under negotiation are included, the total may increase by up to 150 billion US dollars. The "450 billion" in the original title refers to the individual cloud computing commitment, not all computing power-related commitments, which needs to be distinguished when reading.

On the eve of the sprint for a trillion-dollar valuation IPO, this prospectus is not only a financial disclosure, but also an extreme test of the underlying logic of the entire AI large model business model. It raises a sharp question to investors: Is this a technological revolution leading to AGI, or an asset-heavy drudgery deeply bound by computing power infrastructure vendors?

"Accounting factors" in the 80.7 billion loss and the cost structure of the first half of the year

First of all, we must clear the fog of the 80.7 billion US dollars net loss. In normal business logic, a company with half-year revenue of 16.3 billion US dollars that loses 80.7 billion US dollars would have long been insolvent. The reason why Anthropic can maintain operations is that the vast majority of this huge loss comes from non-cash accounting expenses. The 72 billion US dollars figure is large, but it reflects the sharp fluctuation in the valuation of previous financing instruments, not the cash burning rate at the operating level.

However, after stripping out the accounting factors, the operating situation of Anthropic in the first half of the year is still severe. The expenditure on computing power and infrastructure of nearly 9 billion US dollars in the first half of 2026 accounts for about 55% of the revenue in the same period. For comparison, the gross profit margin of mature software companies is usually above 70% to 80%. Under the caliber of the first half of the year, for every 1 dollar of Anthropic's revenue, more than 0.5 dollars must be immediately invested in computing power acquisition. This is no longer the light-asset software logic of "writing a set of code that can be copied infinitely", but closer to a heavy industrial consumption war.

However, a key time difference must be supplemented here: 55% is the static caliber of the first half of the year, which does not represent the company's current ongoing operation status. The company later disclosed to investors that it had achieved adjusted profitability in the third quarter, with an adjusted gross profit margin of about 85%, every 1 dollar of computing power expenditure corresponds to about 1.70 dollars of revenue, and free cash flow has also turned positive. These quarterly indicators are not fully comparable to the prospectus data in the first half of the year, and the specific definition of "adjusted" remains to be clarified before the full audit data is released. But they at least show that there are signs that economies of scale can bring diminishing marginal costs. The problem is whether this improvement can be sustained and whether it can cover the rigid computing power bills for the next ten years.

450 billion long-term commitments: cloud giants are both shareholders and suppliers

The most rigid burden in this prospectus is the approximately 450 billion US dollars cloud computing commitment. The disclosed share allocation is clear: about 111.1 billion US dollars for Google Cloud, about 110 billion US dollars for AWS, and about 31.4 billion US dollars for Microsoft, totaling about 252.5 billion US dollars, with a term of 7 to 10 years, payable "whether used or not", and most of them are non-cancellable. The rest comes from other agreements.

Anthropic did not choose to build large-scale supercomputing centers on its own like Elon Musk's xAI, but deeply bind cloud vendors. Amazon and Google are not only important shareholders of Anthropic, but also important computing power suppliers. This structure of "shareholder and supplier" objectively forms a mutually reinforcing relationship between capital and procurement. However, public reports do not give a one-to-one quantitative closed loop between "the amount of equity investment from cloud vendors" and "the amount of long-term computing power contracts", so it is not appropriate to simply describe it as an accurate circular financing. A more accurate statement is: this is a long-term locked-in infrastructure contract with bargaining power biased towards the supply side. No matter whether the model can be sold or the industry encounters a cold winter, Anthropic must pay for this huge commitment; while cloud giants lock in top model teams through equity investment, obtain long-term revenue through long-term computing power contracts, and retain the underlying entrance for the explosion of AI applications at the same time.

86 billion chip lease: NVIDIA's bargaining position

The up to 86 billion US dollars chip lease commitment reveals another reality: even the world's top AI laboratories still have to accept extremely strong bargaining conditions when facing NVIDIA's computing power supply ecosystem.

The prospectus discloses that Anthropic agrees to rent NVIDIA chips from AI cloud service providers such as Nscale and Lambda, with contract terms expiring in 2033 and 2034 respectively, with a maximum expenditure of 86 billion US dollars, excluding other cloud computing contracts. It needs to be clarified that the direct counterparty of these 86 billion US dollars contracts are "new cloud" service providers such as Nscale and Lambda, not NVIDIA itself; but the underlying computing power leased is NVIDIA GPUs, and the pricing logic is ultimately dominated by NVIDIA's dominant position in the GPU and CUDA ecosystem. Whether buying cards directly or renting them for a long time through cloud service providers, large model companies are essentially paying for the high gross profit margin of computing power suppliers. Some commentators compare NVIDIA's pricing position to a "computing power cartel", but this is only a rhetorical description, not a judgment of its legal status; a more common metaphor is "the central bank in the AI era", pointing to its influence on the supply rhythm of computing power, this key input.

Why must it sprint for IPO? A "game of chicken" that cannot be lost

Understanding these two sky-high commitments, you can understand why Anthropic still pushes forward the IPO even when its valuation expectation is revised down from about 2 trillion US dollars to about 1.5 trillion US dollars. In the face of capital expenditures of tens of billions of dollars and upcoming computing power bills, the capital in the private market is no longer sufficient to cover the costs. At the same time, its competitor OpenAI is also consuming capital at a high speed: OpenAI previously allowed employees to cash out about 6.6 billion US dollars through the old share tender offer, and gave a guidance of 70 billion US dollars ARR by the end of the year; but as of the end of September 2026, its actual ARR is about 50 billion US dollars, and the market still has doubts about the radicalness of the end-of-year target. The two companies are engaged in a race of "who runs out of capital first".

Anthropic's early investors and strategic investors Amazon and Google need an exit channel, and the public market is one of the few places that can accommodate trillion-level valuation expectations. However, against the background of the volatility of US tech stocks and the attention to the debt problem of AI computing power infrastructure, investors will inevitably remain cautious about the more than 600 billion US dollars of infrastructure commitments when facing this prospectus.

Conclusion: Trillion-dollar gamble under the asset-heavy model

In Silicon Valley, the narrative about AI is always full of romanticism. But this prospectus of Anthropic pulls the discussion back to financial reality. At least under the caliber of the first half of the year, large model companies present the characteristics of low gross profit margin, heavy assets and high cash consumption; the improvement of adjusted gross profit margin and cash flow in the third quarter shows that economies of scale may be emerging, but whether it can be sustained and whether it can cover the long-term rigid computing power bills still needs full audit data and time to verify.

For the capital market, this prospectus is a sobering agent. When computing power costs occupy half of the revenue in the first half of the year, and when the 10-year cloud bill locks in future flexibility, AI large model companies must prove that they can generate far higher profit margins than traditional SaaS in the future. Otherwise, the foundation of the trillion-dollar market value will face severe tests.

This article is from the WeChat official account "Tech Business", author: Tech Business, published with authorization from 36Kr.