Zhipu should thank Anthropic

Ask AI · How can Zhipu’s MaaS model learn from Anthropic?

Source | Bohu Finance (bohuFN)

Author | All too well

On March 31, “the first listed company in large models,” Zhipu, released its first annual results since going public.

The financial report shows that in 2025, Zhipu’s total revenue exceeded 700 million yuan, up about 132% year over year. Meanwhile, driven by continued expansion of R&D investment, the company’s net loss for the period widened to 4.7 billion yuan, up 59.5% year over year; adjusted net loss was 3.18 billion yuan, up 29.1% year over year.

But behind the massive losses, Zhipu’s performance in the capital markets has been nothing short of remarkable.

After the financial report was released, Zhipu’s share price surged by more than 30% at one point. On April 1, its market cap crossed the 400 billion HKD threshold.

You can’t help but marvel: the valuation of large-model companies doesn’t rely on the model—it relies on imagination.

01 What is everyone excited about?

Why is the market so excited? Let’s look at this financial report first.

Zhipu’s revenue mainly comes from two lines of business.

The first is localized deployment: for customers with extremely high data security requirements—such as governments, large state-owned enterprises, and financial institutions—the AI system is deployed directly on the customer’s own servers. This “project-based” delivery model has a long implementation cycle and heavy resource input, but its advantage is strong customer stickiness.

In 2025, revenue from this segment grew from 260 million yuan to 530 million yuan, accounting for 74% of total revenue. The growth rate reached 102.3%, making it Zhipu’s current cash cow.

But the ceiling is obvious too.

Project-based operations mean growth is constrained by the sales team’s productivity and the project delivery capacity; renewal rates carry uncertainty. At the same time, implementation costs and resource occupancy are high, so it’s difficult to expand at scale.

A more noticeable change is that the gross margin of this segment has fallen from 66.0% in 2024 to 48.8% in 2025, showing a clear deterioration in profitability. The financial report notes that, to meet customer needs, the localized deployment business投入 more delivery resources, leading to a period-by-period decline in gross margin. Meanwhile, the revenue share of localized deployment is also gradually decreasing.

By contrast, the second line of business—cloud deployment—may be smaller in current scale, but has a larger upside.

Zhipu’s cloud deployment falls under MaaS (Model as a Service, model as a service). It provides model invocation services to developers and enterprises through public cloud APIs, charging by usage. It has a high degree of standardization and lightweight delivery. Growth does not depend on headcount expansion; instead, it depends on model capability and the scale of calls.

In 2025, revenue from this segment was 190 million yuan, up a staggering 292.6% year over year, becoming the fastest-growing business segment for Zhipu.

Even more attention-grabbing is that in February 2026, Zhipu consecutively raised prices for its API services. In the first quarter, the cumulative price increase reached 83%, yet the number of calls grew by 400%—the market still remains in short supply.

Behind this are factors such as a period of tight supply of compute capacity (the Agent scenario explosion brought by OpenClaw). In Agent scenarios, tokens consumed per single task far exceed those of ordinary conversations, and usage frequency is more stable. It also indicates that the model itself truly has strong capability. For example, the GLM-5 released on February 11 scored 50 points in Artificial Analysis’s Intelligence Index at the time, ranking 5th globally and #1 among open-source models.

This is also the core reason why the market is so focused on Zhipu’s cloud business: the path Zhipu is taking has already been validated overseas—and it is exactly what Zhipu is benchmarking against: Anthropic.

Anthropic is a representative case of an API-based commercial model among U.S. AI companies. Its core path is to deliver the strongest model to enterprises and developers through an API.

At present, Anthropic’s enterprise customers exceed 300k. In the past year, the number of customers with annual spending above 100k USD grew by 7x; customers with annual spending above 1 million USD have exceeded 500. In each of the past three years, growth has surpassed 10x annually. It has been called by The Wall Street Journal “the fastest-growing enterprise software company in history.”

After SpaceX filed, Anthropic became the AI unicorn with the second-highest global valuation, only behind OpenAI.

When model capability is strong enough, model-as-a-service is the most core business model.

Zhipu wants to try the salty and sweet.

Among China’s top ten internet companies, 9 have made deep calls to Zhipu’s GLM model. By March 2026, Zhipu’s registered enterprises and users exceeded 4 million, and its service coverage spans more than 218 countries and regions worldwide.

During the annual earnings call, Zhipu CEO Zhang Peng shared that Zhipu’s 2025 core growth engine is the comprehensive breakout of its open platform and API business. Currently, Zhipu’s MaaS platform has about 1.7 billion yuan in annual recurring revenue, which increased 60x over the past 12 months.

Even more worth noting: through engineering optimization on the inference side, Zhipu significantly reduced the token unit cost, and business profitability improved markedly. The MaaS platform’s gross margin increased by nearly 5x to 18.9%, already clearly above the industry average level.

With the boost from Agent (intelligent agent) scenario explosions such as OpenClaw, it’s highly possible that cloud revenue will achieve a real breakout in the next one to two years—this is what excites people.

But things aren’t that simple.

02 The Optimists Move Forward

The question is whether this kind of growth is sustainable.

The financial report shows that a substantial portion of Zhipu’s API revenue comes from China’s large internet firms. Although these companies have self-developed models, in specific business scenarios they choose to call external model capabilities as well. This “multi-model invocation” approach does provide a stable source of demand, but it doesn’t equal truly scalable growth. Relying mainly on top clients—plus fellow competitors—still carries considerable risk.

Zhipu isn’t the only one trying to go down this road.

Compared with MiniMax, Zhipu doesn’t have an advantage in model invocation volume. On the OpenRouter leaderboard, MiniMax topped the list for five consecutive weeks, with monthly call volume as high as 6.9 trillion tokens, while Zhipu and the main models from the Moon/A dark side are both around 2.7 trillion tokens.

At the same time, cloud giants like Alibaba, Baidu, and Tencent are also pushing MaaS. They have far larger cloud infrastructure and more resources of enterprise customers. Competitors like ByteDance and DeepSeek are also closing in step by step on model capability. The battle to attract developers and secure enterprise customers is far from settled.

Corresponding to all of this is R&D cost.

The financial report shows that in 2025 Zhipu’s full-year loss was 8B yuan, up 59.5% year over year. Among that, R&D spending reached 3.18 billion yuan, up 44.9%. This means that for every 1 yuan the company earns, it has to spend 4.4 yuan on R&D.

Meanwhile, capital expenditures fell from 460 million yuan in 2024 to 74.7 million yuan, down 83.8% year over year. In simple terms, Zhipu adjusted its compute procurement from a relatively fixed leasing model to a combination of leasing and service procurement, thereby lowering capital expenditures.

But improvements in model capability still require continuously increasing R&D investment. R&D and compute costs do not naturally decline as the scale of calls expands. On the contrary, the premise for revenue growth itself is pushing up costs.

This leaves large-model companies in a dilemma. To get more calls, you must keep improving model capability. But to improve model capability, you must keep increasing investment.

The faster growth is, the more cost pressure increases.

Of course, this isn’t only Zhipu’s problem—it’s a challenge that the entire large-model industry is facing. Until these issues are truly solved, MaaS can drive growth, but it’s hard to generate profit.

But Anthropic’s growth has been too fast to ignore.

In mid-February 2026, Anthropic announced annualized revenue of about 14 billion USD. Just three weeks later, that figure shot up to nearly 19 billion USD—an increase of nearly 5 billion USD in three weeks. According to internal documents disclosed by The Wall Street Journal, Anthropic, the AI company invested in by Amazon, is expected to achieve its first profit in 2028.

But the problem is: Zhipu is ultimately not Anthropic. The gap between the two is comprehensive.

In terms of revenue scale, Anthropic’s annualized revenue is close to 19 billion USD, while Zhipu’s MaaS platform ARR is about 1.7 billion yuan RMB (about 250 million USD). That’s a difference of nearly 80x. In terms of business model, about 80% of Anthropic’s revenue comes from enterprise-level API calls, while Zhipu’s cloud API revenue accounts for only 26.3% of total revenue. In terms of cost environment, Anthropic benefits from Amazon and Google, with abundant compute supply and lower costs; Zhipu, meanwhile, faces tight compute supply and high costs for adapting to domestically produced chips.

Even so, Zhipu still managed to “ride the coattails” of Anthropic.

But whether it can truly hold up depends on Zhipu’s own real capabilities.

Reference sources:

1. Ahead Lab: Banning, Flattery, and a 380 billion valuation: Anthropic’s 2026 fantastical journey

2. China Entrepreneur Magazine: A lobster became the财神爷 for MiniMax, Moon Dark, and Zhipu

3. Dingjiao One: Zhipu, MiniMax: The more they burn money, the higher the market value

4. AGI Interface: While Zhipu runs like crazy, it’s also bleeding

5. Dolphin Research: Zhipu: Earn 700 million, lose 3.2 billion? Dreams are above; talking about “a small-gr局”

The article cover’s lead image and supporting images are owned by their respective copyright holders. If the copyright holders believe their works should not be viewed by everyone or should not be used for free, please contact us promptly, and this platform will make corrections immediately.

Author statement: Personal views, for reference only