What Anthropic’s IPO Reveals About Pricing Chinese Large Language Models
The article views Anthropic’s post-IPO valuation fluctuations as an example of public markets reassessing the growth quality and commercialization capabilities of large language model companies. On this basis, it analyzes the pricing logic of Chinese large language model companies, arguing that a scarce public float and rapid revenue growth cannot sustain valuations over the long term; future valuations will depend more on whether customers continue using APIs after Token price increases.
Why read this
AI score 88
The article juxtaposes the decline in public-market valuations with domestic Token price increases, providing an analytical framework for comparing the commercialization quality and pricing power of Chinese large language model companies.
Topics
Body · AI translation
Anthropic’s IPO looks more like a stress test for the entire AI industry.
By Wei Linhua
Edited by Liu Junhong
At 11:48 a.m. Eastern Time on November 19, Anthropic (ticker: ANT) officially began trading on the Nasdaq, carrying the halo of the “largest IPO in human history.”
Compared with Anthropic’s finalized offering price of $83, the stock opened sharply higher at $111.50, a surge of 34.3%, and at one point climbed to an intraday high of $115.10. On a fully diluted basis, Anthropic’s valuation peaked at $2.02 trillion, surpassing SpaceX, which went public in June this year, in one fell swoop.
However, the post-opening celebration lasted less than half an hour. By the close, Anthropic’s market capitalization had been on a roller-coaster ride.
Beginning in the afternoon, Anthropic’s share price plunged, falling as low as $92.72 intraday before closing at $97.17. Although that was still a gain of 17.07% from the $83 offering price, Anthropic had shed about $324.4 billion in market capitalization from its intraday high by the close. On a fully diluted basis, the company’s valuation fell from $2 trillion to about $1.76 trillion.
As the bell rang across the ocean, nearly every large-model founder, investor, and secondary-market analyst in China’s AI community stayed up late watching the market. They were not watching the rise and fall of Anthropic alone, but searching for an answer: Now that the “anchor” of global AI valuations has entered the deep waters of the public market, how exactly should Chinese AI models be priced next?
After all, on the eve of the bell-ringing ceremony, Anthropic’s valuation had already been cut again and again—from the initial $2 trillion valuation assigned by investment institutions, to the reduction during Anthropic’s Q3 roadshow, and then amid developments such as the U.S. midterm elections and Nvidia’s sudden leadership change.
For Anthropic, pressing ahead with an IPO at this juncture looks more like a stress test for the entire AI industry.
Anthropic’s IPO has never been a story about whether it is “expensive.” It is a story about “how to calculate its value.”
And the new yardstick Wall Street is using to scrutinize Anthropic today will be applied, unchanged, to every Chinese large-model company.
Anthropic’s valuation rose
But the excess had already been squeezed out
As a “bellwether” for large-model companies worldwide, Anthropic’s 17% share-price gain is perfectly normal.
Investor enthusiasm remains strong. According to information reported by Bloomberg, subscription orders amounted to approximately six to seven times the number of shares available, driven mainly by large indications of interest from institutional investors.
As a result, Anthropic not only had the confidence to raise its price, but also planned to capitalize on favorable market conditions and raise more money for itself.
By November 18, Anthropic had ultimately raised its price from the $70–$78 range to $83. It also forcibly expanded the number of shares offered by 19%, from the originally planned 825 million to 980 million, raising nearly $20 billion more in cash.
With demand for the shares outstripping supply, Anthropic also had a host of computing-power suppliers backing it:
Just before the listing, a group of technology giants that could be described as a “global computing-power all-star team”—including Nvidia, Google, Microsoft, AMD, Broadcom, Micron, and partners of TSMC—banded together and agreed to subscribe for $26 billion worth of Anthropic shares. Google and Nvidia were the biggest spenders, each putting up $10 billion.
By stepping forward, these computing-power suppliers and cloud providers effectively bound themselves closely to Anthropic while securing computing-power orders in advance.
The $26 billion investment looks more like a “strategic lock-in fee” paid by the giants. As long as Anthropic uses the money to continue its breakneck expansion and train its next-generation models, the funds will ultimately turn into orders and flow back through procurement contracts to Nvidia’s chips, Broadcom’s servers, and the data centers of major cloud providers.
Who, then, are Anthropic’s largest shareholders? The answer is Amazon and Google. According to data disclosed by Anthropic, Amazon’s total stake is 14.7%, followed closely by Google with 13.7%.
Although the company’s market capitalization ultimately settled at around $1.76 trillion, viewed over a longer timeline, the seeds of today’s fluctuations in Anthropic’s market value were actually planted three months before its listing.
Back in August, the market was still enthusiastically calculating how Anthropic could be worth $2 trillion. The pricing logic at the time was blunt and straightforward, mirroring the fervor for large models in China’s private market: As long as the models kept improving and revenue growth continued to soar, the company’s valuation potential was limitless.
But reality poured cold water on that enthusiasm. Beginning in October, Anthropic experienced several “moments” of declining valuation:
First, revenue growth fell short of expectations, prompting investment institutions to reassess its value. During its October roadshow, Anthropic disclosed that its quarter-over-quarter revenue growth had fallen from 144% in Q2 to 47% in Q3. Investment institutions therefore moved first to squeeze out some of the excess, cutting its valuation to $1.5 trillion.
The adjustments above still reflected doubts about Anthropic itself, while the company’s subsequent valuation decline was driven more by broader concerns about the outlook for the AI industry.
After the U.S. midterm elections, pessimism over tighter data center regulation triggered another bloodbath for U.S. AI stocks. On November 6, the Nasdaq fell 5.52%, while the Philadelphia Semiconductor Index plunged about 12%.
In the process, Anthropic’s market capitalization declined as its premium based on future potential evaporated.
Moreover, the structure created by tech giants taking sides and tying companies to computing power has also made investors cautious. After all, of the dazzling financing figures and soaring revenue on the books, how much actually comes from organic market demand, and how much is the product of a circular game in which the giants “invest money with one hand and sell computing power with the other”?
With total committed spending under purchase commitments and lease agreements reaching as high as $900.9 billion, Jensen Huang suddenly resigned after 33 years at the helm of Nvidia. Combined with the new signals sent after CFO Colette Kress succeeded him as CEO, this removed one of the most critical pillars supporting the AI computing-power financing game.
On Nvidia’s earnings call, Kress publicly laid out the new rules: Going forward, Nvidia will scrutinize companies receiving financing support extremely rigorously. Guarantees for data centers will take effect only after the facilities have actually been completed and are ready for use, with absolutely no guarantees provided during construction. If construction of a data center is delayed, Nvidia’s guarantee will cease to exist.
Over the past two years, Nvidia has not merely sold the shovels; it has also served as the “super-guarantor” for the entire buildout of AI infrastructure. As of the end of October this year, Nvidia’s exposure from external guarantees for AI infrastructure financing had snowballed to $11.8 billion.
Previously, the market was willing to assign high valuations to large-model giants like Anthropic that were heavily tied to computing power, largely because it assumed Nvidia was standing behind them and would shoulder the burden if anything went wrong. Now the big backer has suddenly been replaced, and the newcomer’s first move was to withdraw the blank check promising unlimited support. It is difficult for the market not to worry.
Once that credit endorsement is no longer effective, investors have no choice but to reassess whether Anthropic’s future growth expectations can be fulfilled—and to what extent.
The same is true in the Chinese market.
Over the past two years, many domestic large-model companies have supported high valuations by tying themselves to major tech companies’ cloud resources and securing strategic investments. But Anthropic’s experience has already spoiled the ending. Once these companies enter the public market, investors will sooner or later have to strip out the boost provided by tech-giant subsidies from their commercialization performance and evaluate the two separately.
Anthropic’s listing sent a clear signal to Chinese large-model companies: The era when a story of rapid growth along a single dimension could win over the entire secondary market is over.
Watching Anthropic fall from the lofty heights of a $2 trillion valuation back to the reality of $1.5 trillion inevitably raises concerns: Chinese large-model companies have likewise relied on revenue growth to drive up their valuations, so what fate awaits them next?
Valuation Deflation, Demand Frenzy
How Should Chinese Large-Model Companies Be Valued?
Any assessment of Chinese large-model valuations must begin with the changes at Anthropic, because it is not Anthropic that has changed, but the yardstick the capital markets use to price companies.
In fact, compared with Wall Street only now beginning to “squeeze the excess” out of Anthropic’s valuation, the market capitalizations of Chinese large-model companies had already moved to more rational levels.
Take the already-listed Zhipu AI and MiniMax as examples. Even before Anthropic rang the opening bell, the two companies had already demonstrated the process of “valuation deflation” for the entire industry.
When Zhipu AI and MiniMax listed on the Hong Kong Stock Exchange in January this year, both were in the limelight, with Zhipu AI even setting a record by reaching a market capitalization of HK$1 trillion. Why were they so outrageously expensive at the time? It came down to two key phrases: scarcity of supply and a narrative of soaring ARR.
When MiniMax listed, only about 17.07 million of its shares were publicly tradable in Hong Kong, accounting for less than 6% of its total share capital. Yet the offering attracted frenzied subscriptions from 420,000 investors and was oversubscribed by as much as 1,837.17 times. Zhipu AI likewise recorded an oversubscription rate of 1,159.46 times.
But just a few months later, their current valuations tell a very different story: Both companies’ market capitalizations have fallen by more than 80% from their peaks. Business revenue is still soaring, but with the lockup periods for restricted shares nearing expiration, the scarcity premium on their limited free floats vanished in an instant.
Zhipu AI’s Market Capitalization
More worrying are the large-model companies still at the table but not yet listed—the chill from the secondary market is inevitably flowing back into the primary market.
OpenAI, the overseas industry leader, has also recently been pressing ahead intensively with a new funding round to replenish its war chest before Anthropic’s listing. OpenAI had originally aimed to raise $30 billion at a valuation of $1.4 trillion, but investors clearly did not have that large an appetite.
Ultimately, the valuation for this funding round was cut from $1.4 trillion to $1.25 trillion, wiping out $150 billion, or 10.7%, while the initial proceeds also shrank to about $15 billion. Under the capital market’s scrutiny, OpenAI could not avoid a single bit of the discount it deserved.
If even OpenAI must bow its head and proactively accept a discount, then turning back to the domestic market, how much room do China’s still-unlisted large-model companies have to raise their valuations?
Once all the excess has been squeezed out, the value of Chinese large-model companies will no longer depend on grandiose promises, but must instead be tied to their actual ability to generate commercial revenue.
Over the past two years, China’s AI industry has been deeply anxious about commercializing large models: MaaS revenue from business customers has been limited, while the consumer market has yet to produce a truly breakout product.
By 2026, as coding took off and demand for agents materialized, token prices began to rise.
According to the Large-Model Token Service Performance Monitoring Report (2026) released by the China Academy of Information and Communications Technology, the normalized average price in China was RMB 6.8 in the second half of 2025, compared with RMB 13.46 since the beginning of this year—nearly doubling. Overseas prices, however, remain higher than domestic ones, with typical models released since 2026 averaging a normalized price of approximately RMB 117.8.
Why are people still clamoring to use them despite the price increases? Because once large models truly penetrate complex production processes such as software development and enterprise-grade agents, customers no longer want cheap conversational toys; they want productivity tools that can directly do people’s work. As long as a model is smart enough and can reliably complete workflows, enterprises are willing to pay real money for it.
During Nvidia’s earnings call, the interim CEO disclosed a new set of figures: rental prices for B300 chips based on the new-generation Blackwell architecture have soared by more than 60% since their launch. Even older cards based on the previous-generation Hopper architecture are commanding multi-year renewal prices as high as twice the original rate. Whether hyperscale cloud providers or emerging compute-cloud companies, all their computing capacity is running at full load.
Domestic computing capacity also remains in short supply. On November 18, Alibaba CEO Eddie Wu said during an earnings call that as production capacity ramps up, domestically produced computing capacity will continue to account for a larger share of overall capacity and become a major engine supporting cloud revenue growth.
Going forward, as supply-chain constraints persist and competition among major companies for chips intensifies, compute cards on the market will only become scarcer in 2027. As continually rising application demand collides with an even tighter supply of computing capacity, token prices will most likely continue to increase.
Whether it is the free-float game in Hong Kong stocks or valuation surges in the primary market, whenever valuations run ahead of commercial delivery, the secondary market will sooner or later squeeze out the excess in the most brutal way possible.
Players who remain at the table must use their “pricing power” over tokens to prove that they are irreplaceable. The fact that token prices rose in 2026 while usage surged has already shown Chinese large-model companies a path to survival: whoever can make enterprises remain unable to do without its API even after a price increase will have secured genuine pricing power.
What Anthropic has sounded is the death knell for the old pricing era. When the tide recedes, large-model companies caught swimming naked will need to take a fresh, hard look at how much they are actually worth.
AI Trends