Abstract:
Recently, Anthropic disclosed its latest operating data to potential IPO investors in San Francisco. Revenue in the third quarter of 2026 was US$16.8 billion, a 45% quarter-on-quarter increase. The first and second quarters were US$4.7 billion and US$11.6 billion respectively. The company also stated that it had achieved "adjusted profitability" in the third quarter and recorded positive cash flow.
On October 16, four people familiar with the matter broke the news to foreign media that as the cost of capital for the expansion of computing power in the AI industry continues to soar, after seeing the latest financial data, some existing shareholders of Anthropic and potential investors interested in participating in its IPO are re-evaluating the value of this AI unicorn and lowering its listing valuation expectations.
Informed sources said that at an investor exchange meeting held in San Francisco on Wednesday, many participating investors discussed company valuations that fell in the US$1.5 trillion range. This figure has shrunk significantly from the market’s previous IPO valuation of approximately US$2 trillion in November.
"If in the past three years, global investors have mainly priced model capabilities, user growth and future imagination, then from now on, OpenAI, Anthropic and a group of large Chinese model companies must answer a simpler question: Can these extremely expensive models form a good business?" Fang Haisheng, founder of Infinite Star, told a reporter from the Shanghai Stock Exchange.
"Close" to making money but the growth rate is declining
In 2025, Anthropic’s annual revenue will only be approximately US$4.6 billion, with an operating loss of US$8.06 billion, of which US$7.33 billion will be spent on computing power and infrastructure. Today, its revenue in one quarter has reached 3.6 times that of last year.
Although revenue continues to grow, the growth rate has slowed down from the beginning of the year. Its revenue in the second quarter more than doubled from the first quarter, while the quarter-on-quarter growth in the third quarter dropped to about 46%.
Previously, analysts who met with Anthropic's management estimated that the company's annualized revenue run rate could reach $110 billion by the end of this year. At present, this goal may not be achieved.
Even more noteworthy is the profit.
The large model industry has always had the business paradox of “the more you sell, the more you lose”. The stronger the model, the more GPUs, data centers, and electricity are consumed for training and inference, and revenue growth may not outpace the cost of computing power. Anthropic disclosed adjusted profits in the third quarter, which means that at least under the current operating standards, scale effects are beginning to appear.
Although adjusted profits and positive cash flow show that commercialization capabilities have been further enhanced, they are not enough to prove that the company has achieved sustained U.S. Generally Accepted Accounting Principles (GAAP) net profits.
At the same time, the capital market has also given extremely high prices. In May this year, Anthropic completed US$65 billion in financing, with a post-money valuation of US$965 billion. Based on a simple annualization of US$16.8 billion in revenue in the third quarter, its revenue scale is approximately US$67.2 billion, corresponding to a valuation of approximately 14 times annualized revenue; if the IPO is ultimately priced at the current market expected valuation of more than US$2 trillion, the multiple will rise to approximately 30 times.
A few days ago, Aswath Damodaran, a professor at New York University’s Stern School of Business, specifically conducted a reverse valuation on Anthropic. He assumes that Anthropic can achieve an after-tax operating profit margin of 30% and a capital cost of about 10% when mature. If the IPO valuation reaches US$2 trillion, then if the company enters the mature stage in 10 years, it will need to achieve annual revenue of close to US$1.2 trillion; if it takes 15 years to mature, the required revenue will be even close to US$2 trillion.
Zhu Xiaohu, managing partner of Jinshajiang Venture Capital, said bluntly: Previously, AI was based on narrative valuation, but in the future it will turn to fundamental valuation.
In addition, the "income" itself of large model companies has begun to require auditing.
Anthropic’s revenue includes sales generated through cloud channels such as AWS and Google Cloud, and the statistical caliber used by OpenAI is not exactly the same. Reuters previously disclosed that about half of Anthropic's revenue last year came from cloud channels, and the company needed to pay about 16% of related revenue to the channel party. Therefore, even if it is called ARR or "annualized income", the actual gold content may be different.
This also means that ARR, which was most often quoted in the primary market in the past, is becoming an indicator that needs to be carefully dismantled.
Cash flow doubtful
OpenAI provides another mirror.
As of the end of September this year, OpenAI’s annualized revenue disclosed to investors was close to US$50 billion. Previously, the market once circulated that it was close to 70 billion US dollars. Later, it was discovered that the difference mainly came from statistical caliber. At the same time, OpenAI was valued at US$852 billion when it completed financing in March this year. The latest reports indicate that OpenAI is discussing raising at least US$30 billion at a valuation of approximately US$1.4 trillion.
However, according to the Financial Times citing internal company forecasts and Reuters reporting, OpenAI predicts that cumulative negative free cash flow may reach US$278 billion from 2026 to 2030, and computing and infrastructure spending during the same period is expected to be approximately US$856 billion.
In other words, even if the revenue has reached the level of traditional large software companies, cutting-edge model companies may still be a giant "capital throughput machine" in the long term.
At present, the capital market has gradually become "desensitized" to whether the revenue of large model companies can be doubled. The next step to verify is how much revenue each dollar of computing power can generate, whether gross profit margins can continue to increase, and how much cash can be left behind due to rapid growth.
China's large model enters financial competition
As of now, the financial picture of China's large model companies is more like that of their American counterparts in their early stages, but their growth is equally rapid.
Zhipu’s revenue in the first half of 2026 was 954 million yuan, a year-on-year increase of nearly 400%, of which open platform and API revenue was 825 million yuan, a year-on-year increase of more than 27 times; the net loss during the same period was still more than 2 billion yuan, and R&D investment was 2.13 billion yuan. As revenue quickly shifted to the API business, the overall gross profit margin dropped to 26.4% from 50% in the same period last year.
The capital market’s re-pricing is extremely “tragic”. As of October 14, the total market value of Zhipu Hong Kong stocks was approximately HK$279 billion, and its market value once exceeded HK$1 trillion in June this year. Model capabilities did not suddenly disappear, what disappeared was part of the valuation premium that the market was once willing to give it.
The situation is similar with MiniMax. The company's revenue in the first half of the year was US$117 million, a year-on-year increase of 283%, and the gross profit margin rose from 12.1% to 17.9%; however, the adjusted net loss expanded to US$293 million, and R&D expenses were close to US$300 million. On October 14, the company’s market value was only about HK$71.3 billion.
It can be seen that the domestic capital market no longer only looks at the model rankings.
Interestingly, the valuations of large model companies that have not yet been listed are more aggressive.
According to reports, Dark Side of the Moon’s ARR has exceeded US$1 billion in August, and its ongoing pre-IPO financing valuation is approximately US$50 billion, equivalent to approximately 50 times ARR; its goal is to further achieve annualized revenue of US$2 billion by the end of the year.
According to the latest disclosure from a third party, DeepSeek’s current annual revenue is also about US$1 billion, and it is currently promoting a new round of financing with a target valuation of 500 billion yuan. Based on current revenue, the valuation is approximately 75 times ARR. It is worth noting that part of its revenue growth comes from model price increases, and the company still uses more than 70% of its computing power for training rather than inference, which means that the competition for capital between commercialization and cutting-edge research and development continues.
Two sets of price systems are emerging for large model companies in China. The public market has begun to use financial reports to correct expectations quarter by quarter. Slight changes in revenue growth, gross profit margin and losses will be quickly reflected in the stock price; while the primary market’s pricing of leading model companies is still more betting on the future, and this valuation often comes from the target price of the latest round of financing. The growing difference between the two is itself a signal that AI valuation has entered the next stage.
In the past three years, the most important question in the large model industry has been "whose model is stronger?" In the next few years, this question will gradually become: Who can sell smart phones more expensively while keeping computing power costs lower.
Comments