Anthropic, powering the second half of the AI era
In the second quarter of 2026, Anthropic delivered a performance report that made the market recalibrate its expected growth rhythm.
Preliminary revenue exceeded 11.5 billion US dollars, representing a year-on-year increase of approximately 14 times compared to 787 million US dollars in the same period last year. The quarter-on-quarter growth rate reached at least 143% — directly jumping to the three-digit level from 4.73 billion US dollars in Q1.
More critically, the company recorded positive adjusted operating profit for the first time in the same quarter, with an operating margin of around 5%.
A 5% margin is not impressive in the traditional software industry, but in a sector that was considered "unprofitable until at least 2028" just a year ago, this figure carries entirely different implications.
It has moved the profitability expectation timeline of the entire AI track forward by two years.
01
If you shift your focus from Anthropic and look at the revenue curves of major global AI companies, you will find a fact: No matter what business model you choose, what market path you take, or whether you use closed-source or open-source solutions, revenue is accelerating.
In the first echelon of the US, OpenAI's annualized revenue rose from 25 billion US dollars in February to over 40 billion US dollars in August, surging by 60% in half a year.
For the full year 2025, OpenAI's actual revenue was approximately 13 billion US dollars, and its internal target for 2026 is 30 billion US dollars. Its growth engine has expanded from the single ChatGPT consumer subscription to enterprise Agent products ChatGPT Work and programming tool Codex.
OpenAI's CFO stated at an internal meeting in July that the monthly incremental ARR has exceeded the total of the entire second quarter — a quarter that was already quite remarkable on its own.
Google's growth curve is even steeper.
After Google Cloud posted 20 billion US dollars in revenue in Q1, representing a 63% year-on-year increase, its growth further accelerated in Q2 to 24.77 billion US dollars, up 82% year-on-year. Among this, revenue from generative AI products saw a year-on-year increase of nearly 800%.
The backlog of cloud business orders continued to expand from 460 billion US dollars in the first quarter to 514 billion US dollars, and the operating margin rose from 17.8% in the same period last year to 32.9%. Enterprise AI solutions have for the first time become the core growth driver of Google Cloud.
xAI in the second echelon is much smaller in scale, but its growth rate is equally staggering.
Grok's full-year revenue in 2025 was approximately 350 million US dollars, and its 2026 target is 2 billion US dollars, a 4-fold jump. The price of this jump, however, was an operating loss of 6.4 billion US dollars in 2025 and infrastructure spending of about 1 billion US dollars per month.
On the Chinese AI side, although the scale is different, the momentum is equally strong.
The annualized revenue of ByteDance's large model business has reached 4 billion US dollars, ranking first in China, but this figure is less than half of Anthropic's single-quarter revenue.
Alibaba Cloud's AI-related products have an annualized revenue of approximately 5.2 billion US dollars, maintaining three-digit growth for 11 consecutive quarters, accounting for 30% of the cloud's external revenue.
Baidu's AI business recorded 13.6 billion RMB in revenue in Q1, up 49% year-on-year, surpassing for the first time the online marketing business that has supported its performance for more than ten years, with its proportion in core revenue exceeding 50%.
In addition, DeepSeek's annualized revenue is 400 million to 500 million US dollars, of which about 90% comes from APIs. Zhipu AI's annualized revenue is around 1 billion US dollars, all from APIs and programming packages, without any C-end products. Moonshot AI's annualized revenue is about 200 million to 300 million US dollars, with APIs accounting for more than 70%.
Clearly, at present, all paths of global large AI models, from enterprise APIs to consumer subscriptions, from closed-source to open-source, are rapidly delivering revenue acceleration curves.
02
Acceleration after the penetration rate crosses the critical point is usually systematic.
However, while acceleration is a common feature, the paths to deliver results are completely different.
Chinese and American AI companies have developed two business models with very different structures.
The US path can be summarized as "enterprise priority, high price and high stickiness".
About 85% of Anthropic's revenue comes from enterprise APIs and developers, generating an average monthly revenue of 211 US dollars per user — more than 8 times that of OpenAI. It has over 300,000 enterprise clients, of which more than 1,000 pay over 1 million US dollars per year. 8 of the top 10 Fortune companies are using Claude.
This revenue structure means that clients are not "trying out" AI, but embedding AI into their core workflows, resulting in extremely high migration costs.
Among all products, programming tools are the sharpest wedge in this path.
Claude Code was officially commercialized in May 2025, reached 1 billion US dollars in ARR in 6 months, and hit 2.5 billion US dollars in February 2026.
Around 4% of all public code submissions on GitHub worldwide are completed by it, doubling within one month. After Shopify engineers adopted it, the time to launch new features was shortened from 24 days to 5 days.
OpenAI's Codex is catching up, and Bret Taylor admitted at an internal meeting that the programming market was once lagging behind Anthropic, but expressed encouragement about its catching-up momentum.
The significance of programming tools goes far beyond revenue — it turns AI from an "occasional use" tool into the default tool that engineers open every day. Once the usage frequency rises, user stickiness is locked in.
The Chinese path takes a different approach.
Even Doubao, which has 200 million daily active users, generates about 90% of its revenue from the B-end. Volcano Engine accounts for 49.5% of China's public cloud MaaS market, with daily Token call volume exceeding 180 trillion.
On the C-end of Doubao, daily revenue is less than 1 million RMB, mainly from e-commerce commissions embedded in the Douyin Mall, while the daily computing power cost of the application reaches tens of millions of RMB.
About 90% of DeepSeek's revenue comes from APIs, and its C-end products are completely free. Zhipu AI's main revenue comes from APIs and programming packages, with basically no C-end products included.
This is a team that sells shovels for the B-end. The shovels sell well, but they are low-priced.
Chinese open-source models account for 41% of global downloads on Hugging Face, surpassing the US for the first time.
In February, the weekly call volume of Chinese models on OpenRouter surpassed that of US models for the first time. The top 6 models in the global mainstream large model call list are all from Chinese teams, with costs only a small fraction of Western flagship products.
In addition, the difference does not lie in "who is more correct", but in that their respective market structures determine the monetization methods.
US enterprises have abundant IT budgets and high acceptance of security and compliance premiums, so the high-price deep-embedding model is feasible.
In the Chinese market, C-end users have low willingness to pay, and B-end users are price-sensitive, so the path of open-source expansion plus API monetization is adopted.
But both point to the same industrial fact: AI has evolved from "something to try" into an integral part of production tools.
Once this transformation takes place, it is difficult to reverse.
03
Revenue is rising across the board, but profitability has started to diverge, which is the most worthy part to dig into.
Anthropic recorded its first single-quarter profit in Q2, with an operating margin of around 5%. In contrast, OpenAI's operating loss rate in the same period was as high as 122% — it burned 3.7 billion US dollars in cash in Q1, and the full-year burn is expected to be around 27 billion US dollars.
xAI posted an operating loss of 6.4 billion US dollars in 2025.
Doubao of ByteDance has an average daily computing power cost of tens of millions of RMB on the C-end, and Guolian Minsheng Securities estimates that its full-year loss will approach 100 billion RMB.
None of the leading Chinese AI companies have publicly claimed to be close to profitability so far.
Both are developing large models, so why have some already started earning structural profits, while others are still exchanging huge losses for scale?
Breaking down Anthropic's source of profit, it does not come from "cost-cutting", but is the combined result of high revenue quality and optimized cost structure:
Inference gross margin rose from 38% to 70% and then to 85%. The computing power cost required to earn 1 US dollar dropped from 0.71 US dollar in Q1 to approximately 0.56 US dollar in Q2. The cache hit rate exceeds 90%. The actual mixed cost of Opus 4.7 is only one-fifth of its listed price. The training cost is about a quarter of that of OpenAI.
On the other side is revenue quality.
About 85% of Anthropic's revenue comes from enterprise APIs and developers — this is high gross margin revenue, billed by tokens, where clients pay for actual usage.
Around 60% to 70% of OpenAI's revenue comes from ChatGPT personal subscriptions, with a monthly fee starting at 20 US dollars.
The marginal cost of the subscription model rises with usage, and heavy users can easily consume the subscription fees of multiple light users. These are two completely different unit economics models.
Google Cloud's 32.9% operating margin is another reference, but that is the overall level of the cloud platform, not just the model layer.
Google's advantage is that it does not only sell model APIs, but a bundle of chips (TPU), infrastructure, Workspace suite and Gemini Enterprise. The full-stack layout allows it to monetize in multiple links.
However, on the other hand, the cost of training the next generation of cutting-edge models is still rising exponentially.
Anthropic signed computing power procurement commitments of over 300 billion US dollars in 4 months, spending more than 160 million US dollars on computing power every day. It earned 11.5 billion US dollars in a single quarter, while its computing power expenditure alone in the same period reached the order of 14.4 billion US dollars.
The sustainability of profitability depends on whether the speed of inference efficiency improvement can continue to outpace the rising speed of training costs. The sample size is too small to be fully certain, but at least, Anthropic has provided a trackable gross margin improvement curve, which no other company has yet achieved.
Now, the question for investors is no longer "how much has revenue grown", but "how much computing power is burned to earn every dollar".
So far, only one company has given a positive answer to this question.
04
12 months ago, the market was pricing "the money AI may earn in the future", which was belief-driven.
Now, with actual revenue and profit data available, the market has begun to anchor valuations with revenue multiples and profit margins.
Shifting from "believing in the potential" to "being able to calculate the returns" is a fundamental shift taking place in the valuation logic.
Anthropic's latest valuation is 965 billion US dollars, corresponding to an annualized revenue of about 47 billion US dollars, with a price-to-sales ratio of about 20x. Its IPO target is 2 trillion US dollars, corresponding to the end-of-year ARR of 100 billion to 120 billion US dollars estimated by investors, with a price-to-sales ratio of 16 to 20x.
Some investors believe that a company with an annual growth rate of 800% should get a 30x revenue multiple even under the "most conservative" estimate, implying a valuation of 3 trillion US dollars.
In comparison, Palantir and Nebius have a price-to-sales ratio close to 55x — from this perspective, Anthropic is "cheaper".
OpenAI's valuation of 852 billion US dollars corresponds to an ARR of about 40 billion US dollars, which is about 21x.
But OpenAI is still in deep loss, and it is expected to burn 27 billion US dollars in cash for the full year.
When valuing by the same revenue multiple, Anthropic has profits while OpenAI does not.
This difference will be amplified in the public market.
Anthropic targets to go public in October, with underwriters Morgan Stanley, Goldman Sachs and JPMorgan Chase, and it is expected to raise more than 60 billion US dollars.
OpenAI will follow closely, targeting Q4 2026 or 2027.
SpaceX (including xAI) has gone public, and its market value has fallen from its peak of 2.1 trillion US dollars to about 1.49 trillion US dollars, pulling back 50% within six weeks — the public market is still exploring the pricing of such assets.
Once entering the public market, the pricing criteria will change.
The private equity stage can rely on growth rate and narrative, but the public market will start to ask about free cash flow, profit margin trends and customer retention rates. Companies with calculable returns will get a premium, while those with unclear returns will be re-priced downward.
The market no longer gives a premium for "also doing AI", but starts to give a premium for "the money AI actually earns".
05 Epilogue
From the strong performance data of major AI giants and large model companies this year, it can be seen that the AI narrative has fully entered a new stage of systematic delivery.
This means it is no longer a beta story, but a story where differentiation begins.
Accordingly, the core question for investment is also shifting. From "can AI make money" to "who is making money", to "what quality of money is being made", and then to "how long can the profit last".
The next truly worthy metrics to track are not valuation multiples, but three indicators: how much computing power cost is needed to earn 1 US dollar, enterprise customer retention rate, and the slope of the inference cost decline curve.
These three figures determine which companies get structural premiums, and which ones only get cyclical illusions.
This is good news for leading companies — valuations supported by real data are more resilient than pure narrative-driven valuations. But for the second echelon and long-tail companies, once the growth rate slows down or profitability fails to be delivered for a long time, the correction will be very drastic.
When a company earns more money in one quarter than the sum of its profits in the past three years, what the market needs to do is not to cheer, but to start calculating the numbers seriously. This stage has only just begun.
Gelonghui Statement: The views in the article are all from the original author and do not represent the views and positions of Gelonghui. Special reminder, investment decisions need to be based on independent thinking. The content of this article is for reference only, and does not constitute any practical operation suggestions. Trading risks are borne by the bearer.
This article is from WeChat Official Account "Gelonghui APP" (ID: hkg