The so - called "Chinese version of Anthropic" is a false proposition.
You must have seen that news this morning, right?
Anthropic has raised funds. It's the Series H round, with $65 billion in financing, and the post - investment valuation is $965 billion. It's just a small amount short of joining the trillion - dollar club.
They also launched a new model, Opus 4.8. I tried it this morning, and the throughput efficiency is indeed much faster. The inference intensity now has four levels, and I haven't fully grasped it yet.
$965 billion. A company founded five years ago is almost valued at $1 trillion.
The answer seems simple: the model is strong. Claude is truly powerful. Its programming ability has long topped the charts, and the performance of the Opus series in complex reasoning is obvious to all.
But this answer is correct but incomplete.
Anthropic is not the only company with a strong model. Google's Gemini, Meta's Llama, and OpenAI's GPT are all excellent. Relying solely on model capabilities can't support a trillion - dollar valuation.
I looked into what Anthropic has done in the past year, and there is a very clear product line.
In May 2025, Claude Code was officially released. It's a command - line programming tool that allows developers to collaborate with AI directly in the terminal to write code. By February 2026, the annualized revenue of this tool had reached $2.5 billion.
Then came Cowork.
It was launched in January 2026. It no longer only serves programmers; non - technical personnel can also use it. It can directly generate tables from screenshots and automatically turn notes into reports. More interestingly, the entire product was developed in just ten days, written by Claude Code itself.
After that came industry plugins.
Legal, finance, and sales plugins were added one by one. Contract review, NDA classification, and briefing writing, which originally required purchasing a whole set of SaaS software to complete, can now be done with just one plugin.
Looking at these three steps together, the path is clear:
First, attract developers with the programming tool, then bring in non - technical users with Cowork, and finally take over the core workflows of enterprises with industry plugins.
This route targets the wallet of the enterprise software market.
I checked, and the market's reaction was quite straightforward. On the day Cowork plugins were released, the global software stocks lost about $300 billion in market value.
Thomson Reuters dropped 16% in a single day, and LegalZoom dropped 20%. JPMorgan Chase issued a report saying, in essence, that Anthropic is devouring everything, and the SaaS business model has nowhere to hide.
From the end of January to mid - February, the S&P North American Software Index dropped to a forward price - to - earnings ratio of 20 times, which is a historical low. The long - term average is 34 times. Within two months, the entire software industry lost about $1.6 trillion in market value in total.
I calculated that Anthropic's current annualized revenue is $47 billion. It is expected to exceed $50 billion by the end of next month. Based on the $965 billion valuation, the price - to - sales ratio is about 18 to 19 times.
Is this multiple expensive?
If Anthropic is regarded as a "model - selling company", a 18 - times PS ratio is indeed not cheap.
Looking at it from another perspective, American enterprises spend hundreds of billions of dollars on software every year. What Anthropic is doing is redirecting the budgets that originally went to companies like Salesforce, Thomson Reuters, and ServiceNow to itself.
In my opinion, the $965 billion valuation is for the harvesting of the trillion - dollar enterprise software budget in the United States. The evaporated $1.6 trillion in SaaS market value is flowing into Anthropic's account.
......
This story spread quickly to China.
In the past few months, "China's version of Anthropic" has become the hottest narrative in the primary market. Investors are asking, founders are talking, and the media is writing.
The question everyone is discussing is actually just one: Who is China's version of Anthropic?
Zhipu is the most obvious one. At the 2025 annual performance briefing, this company, which had always talked about being "China's version of OpenAI", suddenly changed its tune and said it would benchmark against Anthropic.
The logic is not complicated: Anthropic makes money quietly by selling tokens through APIs, and Zhipu also wants to take this path.
MiniMax has been labeled by institutions and the media as "the most similar to Anthropic". The reason is that it is technology - driven, doesn't chase consumer - end traffic, and focuses on APIs and enterprise services. Yan Junjie himself has said that after some hesitation, he chose the technology - driven path, knowing that he would lose something.
Kimi is the fastest - growing. Its ARR exceeded $100 million in early March this year and more than $200 million in April. The K2.6 model has enhanced programming and Agent capabilities and supports the collaboration of 300 sub - Agents.
It seems that the situation is excellent at first glance, but when you look deeper, it's different.
I checked, and the current pricing ecosystem in the domestic AI programming track is quite interesting. Zhipu's GLM Coding Plan is priced at 49 yuan per month, MiniMax Starter is 29 yuan, and ByteDance's Volcengine and Alibaba Cloud's Bailing are offering a special first - month price, as low as 7.9 yuan.
7.9 yuan. The price of a cup of milk tea can get you an AI programming assistant for a month.
What's more interesting is that there is already a dedicated price - comparison website in this track. codingplan.org lists the packages of five or six platforms together to help developers choose the cheapest one.
For comparison:
The subscription price of Claude Pro in the United States is $20 per month, about 145 yuan. Chinese developers pay about one - fifth to one - twentieth of what American developers pay for similar tools.
What's the result of the price war? Tokens are becoming standardized products, and models are becoming SKUs on the shelf. Developers don't care about brands; they care about prices.
The growth data is also worth taking a closer look at.
The ARR of Zhipu's MaaS business has increased 60 times in the past 12 months, even higher than Anthropic's growth rate during the same period. Kimi's growth curve is also very steep. These numbers do make a good story.
I calculated the absolute values.
Zhipu's MaaS ARR is about $250 million. Kimi's has just exceeded $200 million. Anthropic's is $47 billion. The gap is about 200 times. The 60 - fold growth rate is impressive, but the 200 - fold absolute gap is also a harsh reality.
......
200 times. This number is worth thinking about.
In terms of model capabilities, the gap between China's leading large models and Claude is about one to one and a half steps. In programming, reasoning, and long - text processing, some scenarios are already quite close, while others still need one more version of iteration.
This gap is narrowing, and it's not narrowing slowly.
But the revenue gap is another matter. A one - and - a - half - step capability gap corresponds to a 200 - fold revenue gap. This huge gap cannot be explained by "the models are not strong enough".
Where is the gap? Let me list a set of data first:
In 2023, the scale of China's entire SaaS market was 58.1 billion yuan, which is about $8 billion in US dollars. Anthropic's current annualized revenue is $47 billion. The revenue of one company is nearly six times that of China's entire SaaS market.
The statistical scopes are not exactly the same, one is the total industry volume, and the other is the revenue of a single company. What this set of data shows is the difference in the magnitudes of the markets.
It shows a very basic fact: as long as Anthropic can prove that "using me is more cost - effective than using Salesforce", the money will flow to it.
In China, such a market doesn't exist in the first place.
Here is a set of data. The median gross profit margin of Chinese A - share software companies is about 50%. That of their American counterparts is 74% - 75%, a difference of nearly 25 percentage points.
The reason is not complicated. There are still a large number of Chinese software companies doing project - based delivery, which is essentially "selling manpower". They charge based on the number of people.
Chinese enterprises' thinking when buying software is in the same vein. Chinese enterprises' way of buying software is completely different from that of American enterprises. The core is this: American enterprises buy subscriptions, while Chinese enterprises buy deployments.
American enterprises are used to subscribing by seat.
How much does a Salesforce account cost per month, and how much does a Slack seat cost per year? The IT department has a clear software budget item. When an AI tool comes in and replaces a SaaS subscription, the budget is transferred from the old software to the new tool. The path is very smooth.
What about Chinese enterprises? They prefer customized deployments.
They buy a set of systems and install them on their own servers to keep the data in - house. Especially for central state - owned enterprises, local state - owned enterprises, and large private enterprises, the bidding process, data security requirements, and localized delivery make it difficult for the API model based on token - based billing to work.
This is a problem on the supply side, and there are also issues on the demand side.
Let's go back to the growth path of Claude Code. Its flywheel rotates like this: individual developers use it first and pay $20 per month. After getting used to it, they bring it into their teams. When the team starts using it, it triggers enterprise procurement. From the consumer end to the business end, it penetrates from the bottom up.
The first step of this flywheel is "individual developers are willing to pay $20 per month".
What about in China?
The Coding Plan costs as low as 7.9 yuan per month. The entire ecosystem is competing on who can be cheaper. Developers have formed an expectation that AI programming tools should be priced like this.
I checked. Claude Code's global market share in the code generation field exceeds 42%, more than twice that of OpenAI. Cursor and GitHub Copilot are its two largest channel customers, and these two alone contribute $1.4 billion in revenue to Anthropic.
This kind of pricing power doesn't exist in China. Tokens have become standardized products, models are on price - comparison websites, and developers use the cheapest one. The first step of the flywheel fails, and the subsequent enterprise procurement cannot be promoted.
Putting all these together, the 200 - fold gap becomes clear.
The one - and - a - half - step capability gap can be narrowed, but the 200 - fold revenue gap cannot be closed by just improving the models. The essence of this gap is not a technical debt. Chinese large - model companies lack an infrastructure to turn model capabilities into revenue.
The United States has a pool of hundreds of billions of dollars in enterprise software budgets and a payment habit of subscribing by seat. China doesn't have these.
In my opinion, the 200 - fold revenue gap points to a more fundamental question: If the American path doesn't work, where will the money for Chinese AI come from?
......
Let me share my observations first. The answer may not be "Who is China's version of Anthropic".
I remember something around 2013.
When mobile payment exploded in China, many people were asking a similar question: Who is China's version of Visa? Later, we all know the answer. There was no China's version of Visa. Instead, Alipay and WeChat Pay emerged.
What they did was completely different from what Visa did.
In the United States, mobile payment is essentially "substitution". People already have credit cards, and Apple Pay just saves you the trouble of taking out your card, but the underlying credit card network is still in operation.
In China, most consumers don't have credit cards at all.
Alipay and WeChat Pay directly built a new network. They reached places where Visa had never been: vegetable markets, roadside stalls, and pancake carts.
Looking back at AI, I think the situation is similar.
What Anthropic is doing in the United States is also essentially "substitution". Enterprises are already using Salesforce, and Claude says it can do better and cheaper, so the budget is transferred. The market is already there.
What about in China? The market doesn't exist. So where will the money for Chinese AI come from? My conclusion is: from "getting enterprises that have never used software before to use AI for the first time".
Two things are happening. First, model companies are gradually becoming "suppliers".
What's the current mainstream way in China? Cloud providers set up a model supermarket and put the products of five or six model companies on it. Developers can choose according to their needs. Some platforms even have an Auto mode that automatically matches the "most cost - effective" one for you.
In this structure, model companies are becoming more and more like chip companies.
Qualcomm makes chips, and consumers buy mobile phones. No one would say "I bought a Qualcomm today". Model companies provide underlying capabilities, and users perceive the products of the platforms.
In the United States, Anthropic is both a chip company and a mobile phone company. It makes models, develops products, and collects money from enterprises. One company earns profits from two levels.
In China, the structure is being stratified. Model - making is separate from product - making and revenue - collecting.
Second, AI is entering Chinese enterprises through the "embedding" path.
Anthropic's approach is to open its own door and penetrate from developers to enterprise workflows. It's different in China. The door is already there.
What do Chinese enterprises' daily work run on? Collaboration platforms, enterprise communication tools, and cloud service consoles. These places have already accumulated organizational relationships, approval processes, customer data, and chat records.
When AI comes in, it grows directly within the existing workflows, emerging from group chats, approval processes, and production scheduling systems.
I noticed that some collaboration platforms are already proposing the concept of "AaaS", Agent as a Service. The core is to let AI run directly in the systems that enterprises are already using.
Feishu's commercialization growth rate in Q1 2026 exceeded expectations, and DingTalk has signed a large number of leading enterprises in the manufacturing and retail industries.
The same logic applies to the industrial scenario. Baidu's Famo's decision - making agent directly cuts into production scheduling and logistics planning. Business experts can adjust plans by talking to the system in natural language without writing code.
Looking at these two things together, the trend is quite clear: the monetization structure of Chinese AI is taking a different shape from that of the United States.
In the United States, a model company makes models, develops products, and collects money from enterprises at the same time. Anthropic is such a model.
In China, the emerging shape is that model companies provide capabilities, and platforms are responsible for reaching enterprises, embedding into workflows, and collecting fees. The two levels are separated, and each makes its own money.
This is why I think the question "Who is China's version of Anthropic" may be wrong from the start.
It's highly likely that there won't be a single company in China that occupies both levels. Model companies and platform companies will both do well, but their business models are different from Anthropic's.
Just as there was no "China's version of Visa" in 2013, the emerging thing has nothing to do with Visa.