Anthropic Races Toward a Trillion-Dollar Valuation While Zhipu Focuses on API Gross Margin: A Comparison of the Commercialization Paths of Chinese and U.S. Large-Model Companies

By | Large Model Home

As Anthropic on the far side of the ocean knocks on Wall Street's door with a valuation expectation in the trillions of dollars, Chinese large-model vendors represented by Zhipu are heads-down in their financial reports, accounting for every cent of the gross margin on their API business.

On October 26, large-model unicorn Anthropic made public its IPO prospectus, preparing to officially list in mid-November. The prospectus laid the company's sprint-to-IPO fundamentals on the table: revenue of $16.29 billion in the first half of this year, with a net loss as high as $80.7 billion. Before that, market discussions of its highest valuation had at one point reached $2 trillion.

Just two months before Anthropic knocked on Wall Street's door, Chinese large-model companies also delivered a report card. At the end of August, Zhipu disclosed its interim results: revenue of 954 million yuan in the first half, a loss of 2.072 billion yuan, and, for the first time, the gross margin of its API business turned from negative to positive.

The large-model industry, it is clear, is staging two completely different narratives: the far side of the ocean is paying for future technological assumptions with trillion-dollar valuations, while domestic vendors are busy using gross margins to prove that large models themselves are a business that can make money sustainably.

In the past, there was often a misunderstanding in the industry that overseas companies relied on technology and capital to tell stories about the future, while Chinese companies were still worrying about commercialization. Yet relying on capital alone will weaken the market's confidence in pricing the "future." At a recent investor day, discussion of Anthropic's valuation had already shrunk from $2 trillion to $1.5 trillion.

"Trillion-dollar valuation" bets on the future, not the present

In the view of Large Model Home, the capital market has never bid on a company's value today, but rather paid for future growth space.

Anthropic's prospectus embodies this logic to the fullest.

On the surface, Anthropic's net loss of $80.7 billion in the first half looks like a "bleeding IPO." But opening up the financials reveals that $72.4 billion of it was a fair-value loss booked on convertible bonds because the company's valuation soared. Looking only at actual business operations, its operating loss in the first half had actually narrowed to $693 million.

What really excites investors is its revenue growth rate. For the full year 2025, Anthropic's revenue was $4.6 billion, up about 12-fold year on year (operating loss widened to $8.06 billion over the same period, with compute and infrastructure spending reaching $7.33 billion). By the first half of this year, its revenue had surged directly to $16.29 billion, about 14 times the same period last year.

Moreover, of the $4.6 billion in 2025 revenue, about $3.8 billion came from pay-as-you-go usage, with fixed subscription revenue at only $789 million. This means enterprise customers are not buying some software outright once, but continuously paying by Token call volume. In the second quarter of this year, Anthropic's single-quarter revenue reached $11.5 billion, overtaking OpenAI's $6.7 billion in the same period for the first time.

This precisely proves that enterprises now recognize the commercial value of large models and are willing to pay for code generation, model calls and complex tasks.

But what props up Anthropic's trillion-dollar valuation is clearly not just the existing tens of billions in revenue, but the "acceleration" of revenue charging upward. And this is precisely the root of Anthropic's valuation plunge.

Financial data shows that Anthropic's second-quarter revenue was $11.6 billion, doubling quarter on quarter; but by the third quarter, revenue was $16.8 billion, with quarter-on-quarter growth slowing to 45%. Although it is still growing rapidly overall, as long as the growth rate falls short of Wall Street's expectations, the valuation logic will immediately waver. Subsequently, discussion of its valuation fell back from $2 trillion to $1.5 trillion.

Merely because of one quarter of slowing growth, $500 billion in valuation expectations evaporated into thin air. This is the price of pricing by "buying the future": what investors pay for is the expectation of growth, and once the slope of that expectation loosens, the price is immediately settled.

The path to commercial monetization for "Chinese models"

"The API business has a fatal risk: it can all be cut off overnight." Zhu Xiaohu, managing partner of GSR Ventures, poured this bucket of cold water when evaluating the domestic large-model industry this year. As an investor who has backed Didi and Ele.me, he has an almost instinctive vigilance about the fragility of business models.

But on paper, the transformation of domestic large-model companies over the past six months has indeed been quite fierce.

In particular, Zhipu, once called the "Chinese version of Anthropic" in the industry, is seeing its anchor positioning effect weaken as Anthropic's IPO approaches, and Zhipu is also exploring its own commercial path. Revenue in the first half of this year was 954 million yuan, up 399.7% year on year, exceeding the whole of last year in just half a year. More important than the growth rate is the complete flip in revenue structure: in the first half, its open platform and API business revenue reached 825 million yuan, surging 2735.7% year on year, with its share of total revenue jumping directly from 15.2% in the same period last year to 86.5%; while localized deployment revenue, which used to carry the load, fell from 84.8% to 13.5%, and enterprise-level general large-model revenue declined 54.6% year on year.

Simply put, Zhipu changed its way of living in half a year: the "selling boxes" business of negotiating projects one by one and doing localized delivery is rapidly shrinking, and cloud services billed by Token call volume have become the absolute mainstay.

More critically, this growth was not achieved through a "price war."

In the first half of the year, Token call volume on Zhipu's MaaS platform grew more than 40-fold compared with the start of the year, paying daily active users increased 603%, while the average API selling price instead rose by about 101%. Zhipu CEO Zhang Peng summed up this commercial logic in a formula: "AGI commercial value = intelligence upper bound × Token consumption scale," in which model capability determines pricing power and call scale determines the base. In the first half of this year, these two curves moved upward in sync for the first time, and also turned Zhipu's API gross margin from negative to positive, reaching 24.6%.

Moving from "doing projects" to "selling APIs" has indeed become the consensus of the domestic industry. Not only Zhipu: MiniMax's revenue in the first half of this year was $117 million, up 283.1% year on year, and the share of revenue from its open platform and enterprise services also rose from 30.3% last year to 63.4%; in August its annualized recurring revenue (ARR) broke through $800 million, 80% of which came from the enterprise side. China's large-model companies are not just one company switching tracks; they are collectively switching business models.

However, as AI technology becomes widespread, APIs will become a public utility like water and electricity, and gross margins will eventually fall back to the 10% to 20% level. Zhu Xiaohu even said bluntly that the reason Anthropic is rushing to go public this year is that it clearly knows that once computing power becomes widespread in the future, the days of earning 60% to 70% gross margins from frontier models and coding tools will simply not be sustainable.

The implication is that the API business can make money today entirely thanks to the technical barriers still held by leading models; it will be hard to make money in the future because those technical barriers will sooner or later be flattened.

The capital market reacted faster than anyone to this endgame risk. On September 10, DeepSeek released the V4.1 Flash model and announced price cuts; that same day Zhipu's stock price still fell 10.34%, and investment bank Jefferies subsequently cut the valuation multiple for Zhipu's API business from 50x to 30x.

Competition in large models shifts from revenue to cost

When it comes to China's large-model market, the most intuitive impression from the outside is outstanding "cost-effectiveness." Data shows that Silicon Data's Token price index fell all the way from $2.04 at the end of May this year to $0.97 in early September, falling below the $1 per million Token threshold for the first time, and the main driver behind this was precisely China's open-source models.

But Morgan Stanley's survey data points out that the actual average price of Chinese large-model APIs has not fallen but has risen: in the second quarter of 2026, the average input price of domestic models was 4.9 yuan per million Tokens and the output price was 21.9 yuan per million Tokens, up about 48% and 80% respectively from the first quarter of 2025. Looking at individual products, the three subscription tiers of Zhipu's Coding Plan were raised by 141%, 261% and 130% respectively; Moonshot AI's Kimi K3 rose more than threefold; even DeepSeek, which has always led price cuts, sharply raised the call price for cache hits during peak hours by 12-fold.

Why is it that open-source models are leading price cuts on one side, while average commercial API prices are surging on the other?

Data from the National Data Administration shows that domestic average daily Token call volume has soared from about 100 billion at the start of 2024 to 140 trillion in March this year. When the pool is deep enough and the application scenarios are complex enough, vendors no longer need to rely on crude one-size-fits-all price cuts to win customers, but have learned peak-and-off-peak pricing, cache pricing and tiered subscriptions.

At the same time, domestic vendors are also using technology to push down underlying inference and training costs. Zhipu's per-unit Token inference cost in the first half fell 80% from the start of the year; a UBS report also estimates that the training cost of China's leading models may be only one-tenth that of overseas peers.

If one goes back at this point to look through Anthropic's prospectus, one finds that in its S-1 filing its computing power procurement commitments for the next ten years amount to as much as $518 billion, of which about 80% is non-cancelable—you have to pay whether you use it or not. In addition, 47% of its revenue last year was completed through Amazon's and Google's cloud marketplaces, and for every $1 received it has to pay cloud vendors a 16-cent distribution fee.

In other words, on the matter of cost structure, the roles of Chinese and American vendors have been reversed. Anthropic, which is seen as "telling a future story," has locked its computing power costs firmly at a high level with a ten-year contract; while Chinese vendors, seen as "struggling to find monetization," have written a continuously declining cost curve through engineering optimization.

Final thoughts

Whether it is a giant sprinting toward a trillion-dollar valuation or a startup carefully calculating gross margins to survive, in the end all must submit their answers within the same commercial loop: technology creates customer value, customers pay to validate that value, and revenue feeds back into next-generation R&D.

Returning to the question at the beginning: who exactly is pricing the large-model industry?

The answer today is increasingly clear: what prices the industry has never been the underwriters in the roadshow hall, but the real daily Token call volume, the quarter-by-quarter changes in gross margin, and the slopes of those two curves—falling costs and rising revenue. No matter how the story is told, when the financial data are published, it will be clear who is swimming naked.