Behind Anthropic's Huge Prospectus Losses: Accounting Losses and Rigid Compute Commitments

By | Tingtong Tech (ID: tingtongtech), Author | Yang Lin, Editor | Rao Xiafei

Put the giants' data side by side, and it forms a set of glaring contrasts for the industry.

On October 26, 2026, Anthropic submitted a prospectus to the SEC, aiming to list on Nasdaq in November at a valuation of $1.5 trillion to $2 trillion.

If priced at the upper limit, this would be one of the largest IPOs in U.S. stock market history. But the prospectus runs headlong into a set of numbers that make you want to rub your eyes: revenue of $16.286 billion in the first half of 2026, up about 1,300% year over year, yet a net loss of as much as $80.748 billion.

Earlier, on October 20, SpaceX released its third-quarter financial report, with revenue of $14.575 billion and net profit of $4.404 billion, achieving its first quarterly profit since going public. And what still props up 85% of the group's operating profit is the Starlink communications business.

On one side, a model company is losing astronomical sums; on the other, the company "selling shovels" has started making money. Behind this is not who loses and who earns, but the cost structure.

For instance, SpaceX's profitability has always relied on a communications business that has passed its peak capital expenditure to backstop the AI segment. Unfortunately, Anthropic has no such buffer: every dollar of revenue must directly face the test of inference costs and rigid compute commitments.

Anthropic's Massive Loss

To understand Anthropic's prospectus, you first have to understand how the company came to be.

In early 2021, Dario Amodei left with his sister Daniela and five other core OpenAI members to found Anthropic. Dario was deeply involved in the development of GPT-2 and GPT-3 and is one of the co-inventors of RLHF.

The reason for leaving was not money, nor position, but an irreconcilable disagreement over OpenAI's direction of accelerating commercialization after Microsoft came on board. Dario increasingly did not believe that the phrase "the stronger the capability, the safer it naturally becomes" would come true on its own.

In May 2021, Anthropic completed a $124 million Series A. But before this money arrived, early fundraising was not smooth; the vast majority of institutions did not invest, and even the Series A funds arrived in three installments.

Anthropic's business logic also took a different path: from day one, it did not choose to compete head-on with ChatGPT in the consumer market.

It placed its chips on enterprise customers and APIs. Data shows that in the first half of 2026, unlike OpenAI, whose revenue mostly comes from consumer subscriptions, Anthropic's API revenue billed by actual token usage was $13.1 billion, accounting for 80% of total revenue; subscription revenue was $3.2 billion, accounting for only 20%.

Its flagship coding tool Claude Code was also a core variable in growth. Just six months after launch, the model reached $1 billion in annualized revenue, and by early 2026 annualized revenue exceeded $2.5 billion, capturing about 54% of the AI coding tool market.

This supported Anthropic's rapid growth. In 2023, the company's first year of commercialization, revenue was $41 million. In the following 2024-2025 period, this figure was $386 million and $4.585 billion, respectively. By the first half of 2026, it jumped directly to $16.286 billion, up about 1,300% year over year.

But the benefits and costs of enterprise customers are two sides of the same coin.

For instance, strong willingness to pay, high switching costs, and predictable revenue are attributes of enterprise customers; the cost is long sales cycles, high customer acquisition costs per account, and the fact that when competitors are willing to knock on the door with discounts, "switching costs" are not as high as imagined.

However, what shocked the market this time was the size of the loss.

Chart: Anthropic's financial situation

Source: Anthropic prospectus, screenshot by Tingtong Tech

The prospectus shows that in the first half of 2026, Anthropic's net loss was $80.748 billion. However, on closer inspection, of this $80.7 billion, $72.388 billion came from fair value changes in convertible notes, derivative liabilities, and SAFE instruments, accounting for 89.6% of the net loss.

The mechanism is not complicated. Because Anthropic went through multiple rounds of financing before going public, its valuation climbed from hundreds of billions of dollars to $965 billion post-money in the Series H round.

Under GAAP requirements, convertible notes and SAFE instruments issued to early investors must be measured at fair value. Each time the company's valuation rises, the liability side appreciates accordingly, showing up as a loss on the income statement.

This is more like a technical product of the collision between accounting standards and the private valuation mechanism, rather than evidence of deteriorating operations. Data shows that for the full year of 2025, the same accounting treatment had already created a paper loss of $34.308 billion, and the first half of 2026 merely continued this logic.

In other words, the more valuable the company, the larger the paper "loss."

The $6.2 billion in convertible bonds held by Amazon is the most typical example. By the end of June 2026, the book value of these notes had been revalued directly from $40 billion at the end of 2025 to $108.566 billion.

Behind the Loss

Setting aside the financial picture, Anthropic's operational problems are much more subtle.

Data shows that in the second quarter of 2026, Anthropic achieved operating profit of $974 million, with the core driver being compute and infrastructure costs, whose share of revenue plunged from 251% in the first quarter of 2025 to 48% in the second quarter of 2026.

Of course, economies of scale are indeed taking effect. As the revenue base jumped from the billions to the tens of billions, the compute resources locked in earlier were spread thinner, and the unit economics improved directionally.

But the price of that improvement shows up on the compute cost sheet.

Data shows that Anthropic's locked-in compute infrastructure commitments have exceeded $600 billion, and if the additional agreements under negotiation are finalized, the total could reach $750 billion to $810 billion.

Among them, the contract with Google Cloud is about $111 billion, with AWS about $110 billion, with Microsoft about $31 billion to $44.9 billion, plus TPU leasing obligations tied to Broadcom of more than $160 billion, and chip leasing arrangements with Nscale and Lambda totaling up to $86 billion.

More critical is the change in contract structure. Many of these agreements are marked as "payable regardless of actual usage," with terms extending to 2033 and 2034.

This also means Anthropic is trading certain spending over the next seven to ten years for certainty of current compute supply. And the valuation of a pure model company still, to a large extent, parasitizes the cash flow of infrastructure suppliers.

Many analyses hold that while revenue growth remains high, this structure works well; but once growth slows, rigid contracts will turn from a scale lever into a profit black hole.

In fact, in the third quarter of 2026, Anthropic's quarter-over-quarter growth was about 45%, a clear brake compared with the doubling or more quarter-over-quarter growth in the first half of the year.

However, the slowdown itself is not surprising; no company on a hundred-billion-level revenue base can sustain triple-digit growth for long. The problem is that Anthropic's contractual obligations will not shrink accordingly as growth slows. This is also the most fragile part of the "trade time for space" strategy: it bets that revenue will run faster than the bills.

Anthropic's highly concentrated revenue structure also sounds an alarm for operations.

Data shows that in the first half of this year, about 20% of Anthropic's revenue came from just two customers, indicating high dependence on cloud platform channels. A concentrated customer base, combined with contractual relationships not yet locked in for the long term, leaves gaps in the quality dimension of Anthropic's high growth that still need to be verified.

Of course, the turning point still appears in model generations.

In late September, the release of Claude Opus 5.5 reduced running costs by 40% compared with the previous generation, and Anthropic's call popularity on aggregation platforms quickly rebounded. This also means that "leading model capability" can still bring pricing power, but the length of this window is being continuously compressed by competitors' discounts and speed of follow-up.

The Rippling case is a slice of this.

In May 2026, the share of developers using Anthropic models for coding work was 57%, while OpenAI had only 16%. A few months later, OpenAI won orders with discounts of 20% to 40% off normal prices plus free credits, and the shares reversed to OpenAI 52% and Anthropic 18%.

And Replit switched its main model to OpenAI after OpenAI offered a plan with an annual value of more than $100 million and discounts of more than 30%.

Industry insiders believe the essence is that Anthropic's pricing strategy is rather tough: once customers hit the contract usage cap, discounts are simply canceled. But when competitors are willing to trade negative margins for share, Anthropic loses its advantage.

Compared with Anthropic's huge losses, SpaceX's third-quarter financial report provides another reference sample.

On October 20, 2026, SpaceX released its third-quarter financial report, with revenue of $14.575 billion and net profit of $4.404 billion, achieving its first quarterly profit since going public.

For example, the communications business where Starlink sits propped up 85% of the group's operating profit; while the superintelligence segment, whose revenue share exceeded half for the first time, had operating profit of only $431 million for the quarter, corresponding to capital expenditure of as much as $16.4 billion.

In other words, SpaceX's AI business is growing fiercely, but its way of making money is selling compute infrastructure, not bearing the marginal cost of model inference.

Starlink's communications business contributes the vast majority of profit precisely because it is a business that has already passed the peak of capital expenditure and whose marginal costs are relatively controllable.

The cost structure facing model companies is completely different.

Many analyses point out that the traditional internet logic of marginal costs trending to zero fails in the face of AI inference. Every user call corresponds to real GPU consumption, electricity spending and infrastructure operations and maintenance. This "diseconomy of scale" will be amplified, not alleviated, during the stage of rapid user growth.

JPMorgan's judgment in this round of downgrades is worth reading closely: it cut the expected price-to-earnings ratio for 2030 of a certain domestic model company from 30 times to 20 times, on the grounds that "the leading position of frontier models is rotating rapidly, and the certainty of a single vendor maintaining technological leadership over the long term is lower than previously expected."

This sentence pierces the fragile link in the valuation of AI model companies: technological leadership is an asset that automatically depreciates over time. When the market is willing to pay a premium for scarcity, model companies' valuations can run detached from financial fundamentals; but when scarcity is diluted by competition, the pricing anchor will quickly fall back.

"Whoever can sell intelligence at a higher price while pushing compute costs lower will have the chance to turn revenue growth into sustainable cash flow," Liao Ge said.

Aswath Damodaran of New York University also provides a quantitative reference: if Anthropic's IPO valuation reaches $2 trillion, the company would need to achieve annual revenue of nearly $1.2 trillion to enter maturity ten years later; if the maturity period is extended to fifteen years, the required revenue would be nearly $2 trillion.

This number itself does not constitute a judgment of right or wrong. But it means the time window left for "burn money first, prove later" is narrowing.

Anthropic's management is clearly aware of this. The turn to positive single-quarter operating profit in the second quarter of this year was mainly to tell the market that model companies can be profitable at the operating level and that economies of scale really exist.

Only, the price of profitability is hundreds of billions of dollars in rigid compute commitments over the next decade.

"In fact, Anthropic's losses, and the fact that most of SpaceX's profit comes from the infrastructure side, both point to the same conclusion: competition in the AI industry is shifting from who tells a better story to who gets their unit economics working first," Liao Ge said frankly.

At the same time, Liao Ge still believes that huge losses themselves are not frightening; the question is, after the accounting components in the losses are stripped out, how long, how much revenue and how high a gross margin will be needed to fill the remaining gap that truly needs to be filled.

Anthropic's prospectus lays this question out on the table. Next, the market may use the same ruler to measure every AI company.

(Statement: This article is for information exchange only and does not constitute any investment advice.)