Who Is Feeding Anthropic: $4.6 Billion in Revenue and Over $8 Billion in Losses

"Claude hunts down Chinese people" is the hottest AI meme on the Chinese internet right now, bar none.

Someone combined Anthropic's various restriction measures into a role-playing game, in which players take on the role of an Anthropic inspector and must pick out the disguised Chinese users from a crowd of Claude users.

The founder's past, and the rumors about Baidu, keep adding a few more layers of intrigue to the picture.

A company like this heading to the capital markets, carrying a massive valuation of $2 trillion and about to sprint toward the largest IPO in history, leaves some rational analysts and the general public confused, drowned in unverifiable rumors and joking memes. Something is off. Or rather, the risk is high.

What kind of company is Anthropic, really? Is Chinese users racking their brains to bypass access restrictions and keep their accounts alive strong evidence for the high valuation?

There is more and more information that adds mystique to Anthropic. To figure out how much it is really worth, and whether that value can last, one needs to peel back the layers like Sherlock Holmes and pull the truly important clues out of the puzzle.

Next, let's arrive together at Anthropic's place of business.

01

Walking into the lobby, who is making all the noise?

Search for Anthropic online, and the first thing you see is all kinds of rumors, which make the company seem especially mysterious. Some people are discussing how to actually get to use Claude, some are exchanging notes on what tricky programming problems CodeX has solved now, and others are asking about accounts and sharing usage tips.

Together, these stories form our first impression of Anthropic. But is that all there is?

Look closely at this group of discussants active on public social platforms, and they roughly fall into three categories:

The first are programmers. They bring tales of the product's capabilities, share how CodeX offers an irreplaceable experience in AI programming, compare the performance of different models, and treat Claude as an indispensable tool in their daily development work. An AI company that the general public cannot easily access and does not know well has somehow captivated some technical users to this degree, and a sense of scarcity is born from that.

That in turn attracts the second category: account brokers. The more Anthropic sets up barriers and the harder it is to use smoothly, the more public curiosity is stirred, and since most people lack programmers' technical skills, some have started businesses selling accounts, acting as purchasing agents, and offering third-party services. Clearly, none of that revenue flows to Anthropic.

What is wrong with Anthropic, that it does not even want to make money in the China region? These strict restrictions, which do not fit the logic of maximizing commercial profit, have sparked all kinds of speculation and attracted a third category of discussants: onlookers.

They are not heavy AI users and may never have actually used Claude. But that does not stop them from joining the discussion. Some link the founder's work experience at Baidu to Anthropic's policies, suggesting the CEO may be suffering from China PTSD.

Others analyze Anthropic's corporate culture, a hard-edged, unsentimental tech-geek temperament that does not care about Chinese users, nor about copyright giants like Disney and Marvel, and generates content with no taboos, a kind of rough-and-tumble "go ahead, try to kill me" attitude.

At this point, we can see clearly that Anthropic's image on the Chinese internet is mainly built by these three groups: a relatively small group of programmers, gray-market operators making money off scarcity and anxiety, and onlookers who do not care about the product itself.

So the people talking the most energetically are not the company's core business, and cannot really explain what kind of company Anthropic is or what its revenue actually looks like. We cannot keep standing in the lobby listening to hearsay; we need to go somewhere closer to the truth: the finance office.

Real money does not lie. Compared with rumors that are hard to tell true from false, the numbers on the financial statements clearly reveal better what kind of business Anthropic is running.

Thanks to its IPO plans, Anthropic has some public data we can use as clues. Open its books, and the first key clue is $4.6 billion in revenue—and who contributed it?

According to Reuters, Anthropic's 2025 revenue was close to $4.6 billion, roughly 12 times growth over the previous year. It is precisely this astonishing speed that supports the extremely high valuation the capital markets have given it. So who contributed the 12-fold increase in revenue?

(Screenshot of the Reuters report)

More than 80% came from usage-based services, that is, charges by API and token. This is generally professional users, such as programmer developers, or continuous calls by enterprises. On the AI business map, this belongs to the B2B market. Less than 20% came from subscription services, that is, monthly and annual membership plans that individual consumers can buy directly in the app.

Now we know that the B2B market of enterprises and developers is Anthropic's main source of revenue. What about the cost side?

The next key clue is that the two major cloud platforms, Amazon and Google, contributed about 47% of Anthropic's sales.

Reuters reports that in 2023, only 11% of Anthropic's revenue was completed through cloud platforms, and by 2025 it was close to half. This is an important indicator: the B2B AI business is highly dependent on customers' existing digital infrastructure.

Anthropic's growth depends not only on how good Claude itself is, but also on whether it can use large cloud platforms to sell its models to more customers. Imagine an enterprise that is already using AWS or Google Cloud. When it wants to integrate AI into its own business systems, purchasing model services through a familiar cloud platform is often easier to fit into the enterprise's existing technical architecture and office workflows, making it the first choice.

On the one hand, it can leverage the sales channels of the two tech giants to reach a large number of enterprise customers. But channels are not a free lunch; Anthropic needs to pay distribution fees and computing costs to these cloud platforms.

In 2025, Anthropic's operating loss exceeded $8 billion, of which computing and infrastructure spending alone reached $7.3 billion, far exceeding revenue. Clearly, it is back to the old script of "the ones selling shovels make the first money."

In addition, these tech companies behind large cloud platforms are often also developing their own AI large models, or supporting R&D by companies they have invested in—for example, Amazon investing in OpenAI, and Google developing Gemini in-house—so they have incentives to keep Anthropic in check.

A small aside: although the Chinese internet has hyped Claude as very expensive, in reality the U.S. market contributed nearly two-thirds of Anthropic's sales. It cannot be ruled out that some of this comes from Chinese users taking roundabout routes, but the proportion is very limited. What truly determines Anthropic's commercial value is the B2B customers who continuously call the models, and the cloud platforms that bring Claude into more business scenarios.

The next question is: how loyal can Anthropic's customers really be? Will they pay for it over the long term?

How can user loyalty be ensured to consolidate revenue? Making the model better than OpenAI's and Google's is a correct but useless statement. In the tech market, there are far too many examples of companies doing well in technology and products but failing commercially.

Serve several top large customers well, embed into customers' workflows, and increase switching costs? This model is simply unsuitable for the overseas IT market, where manpower is scarce and expensive, nor can it reduce marginal costs through scale, and it fundamentally cannot support a high premium.

Which path leads to Rome? The answer is every path, but add a buff to every option.

For example, Anthropic treats technology R&D as the core of commercial competition, but at the same time it places extra emphasis on AI safety. The way it emphasizes this is by creating a crisis that is hard to verify, saying AI is no longer safe, but we can provide protection against it.

For example, in June this year, Anthropic published a grand article saying that Claude had already participated in a large amount of internal code production, that AI was already self-correcting and self-building, and that everyone should slow down AI R&D, otherwise the consequences would be unimaginable. Now four months have passed, and no unimaginable consequences have been found. But this kind of "AI doomsday theory," when it encounters Western societies shaped by Christian culture, can indeed attract many people to pay. Since AI is so dangerous and yet cannot be avoided, then go to the safer Anthropic to take refuge.

The cooperation with cloud platforms does not stop at simple buying and selling; through long-term procurement, it also binds Anthropic's interests together with those of the cloud vendors.

According to Reuters, over the next seven to ten years, regardless of usage, Anthropic will pay at least $111.1 billion to Google, $110 billion to Amazon, and $31.4 billion to Microsoft under long-term infrastructure service obligations.

Clearly, this is an aggressive computing expansion plan that does not match its own business needs. Anthropic is not only betting on business growth, but also building long-term relationships with cloud vendors, thereby reducing competitive friction in the model market.

Locking in computing power in advance will increase Anthropic's financial burden and even widen its losses, but it also forms a stronger alliance between Anthropic and the cloud vendors, because cloud vendors are currently also trapped in the risk of an AI bubble; the two sides share weal and woe.

At this point, an interlocking closed loop has formed: to support model training and long-term competition, Anthropic must obtain enough computing power; to obtain computing power and expand customer coverage, it needs to cooperate with large cloud platforms and bear huge costs; and whether these costs can be sustained depends on whether the product team can turn expensive computing power into tools that users are willing to pay for.

Of course, this cycle can also reverse. If competitors narrow the model gap, customers turn to alternatives, and computing costs cannot be effectively converted into revenue, then Anthropic's huge investment may turn from an advantage into an inescapable financial burden.

04

Outside the building

Along the way, we seem to have uncovered this company's secrets. Revenue is growing fast, but it carries huge infrastructure investment commitments, future costs are enormous, and the competitive landscape is also uncertain.

Anthropic's business situation and logic are not hard to understand, so why is the market full of confusing rumors and stories? Perhaps because only stories can give a commercial company charm. Tech enthusiasts gain hope from the stories, onlookers get excitement from the stories, and early investors cash out money through the stories.

Anthropic itself is not mysterious. Someone has taken a business story we are familiar with, smeared a new layer of mysterious paint on it, and put it up for auction again. And the place where new stories are auctioned is outside the building.

This article comes from the WeChat public account "Naojiti" (ID: unity007), author: Zanghu, published by 36Kr with authorization.