Zhipu: Make 700 million, lose 3.2 billion? The dream is still there, but talking about losses "shows a narrow perspective"

6x in a quarter—this period’s absolute darling of capital, a star stock of the model sector—Zhipu has also turned in its report. Compared with such a soaring share price, the company’s performance in the second half of 2025 looks much calmer—but does that really matter?

Straight to the main course:

I. Revenue side: March API business ARR is $250 million—everything else is minor!

Among several independent model vendors in China, $Zhipu(02513.HK) is essentially a “pure domestic large model” player. Its talent pool is backed by universities, and its customers are largely B-side accounts such as government agencies and state-owned enterprises.

With heavy requirements for on-prem deployment and delivery, the market has long worried about the continuity of renewals for on-prem deployment projects. Therefore, before the Agent boom, the market was not willing to award Zhipu much of a valuation premium.

1) Revenue slows down, but the open-platform business catches up: Total revenue in 2025 is $720 million, up nearly 132% year over year. It’s still surging fast, but compared with last year’s 160% growth, the growth rate has eased.

Since the first-half results have already been released, the key is to look at the second half separately: revenue is $530 million, up only 99% year over year—an even more obvious slowing trend.

And the main reason the earlier capital-market concerns were indeed confirmed is this: the on-prem deployment business, accounting for 80%+ of revenue, slowed to 57% growth in the second half of 2025, reaching $370 million, while its revenue share also slipped to 70%.

In contrast, what the market truly wants to see—the API interface and cloud-deployed open-platform type of business—grew sharply by 430% year over year in the second half, reaching $160 million. The full-year API-type business volume of $190 million is basically comparable to MiniMax’s $180 million. Moreover, looking at the growth slope, Zhipu’s full-year growth of nearly 300% is more explosive than MiniMax’s 200%.

This is also the core reason the market is willing to give Zhipu a higher valuation: the model’s API interface and open-platform business do show weaker gross-margin performance, because the industry’s aggressive competition before the large-model “arms race” led to aggressive pricing for this segment.

But once standardized, light-delivery revenue like “model-as-revenue” is scaled up, the upside is effectively boundless—especially since, in February, the company also raised prices repeatedly for API interfaces and related offerings. In a single quarter, the price increase was 83%, and gross margin importance becomes much less critical.

Revenue growth is the key metric that validates Token consumption and how welcomed the model is: with the open-platform growing so rapidly in the second half even before GLM 5 is released, it does indicate that Zhipu’s model is truly “substantive.”

And since Chinese New Year this year, Zhipu has iterated the model in three quick waves:

1. On February 11, GLM-5 is released. At the time of release, it ranked #1 open-source on Artificial intelligence’s Intelligentization Index leaderboard;

2. On March 15–16, just one month later, for the suddenly popular “Lobster Agent,” it launches a dedicated GLM-5-Turbo model, focusing on agent workflows such as tool calling, multi-step task execution, complex instruction decomposition, and multi-agent collaboration.

3. On March 27, focusing on the programming domain, GLM-5.1 as a post-training optimization version of GLM-5 is opened to all Coding Plan users.

At the same time GLM-5 is released, the company directly announces a large-model price increase—subscription prices and API prices both rise across the board. Among them, the API price rises again after Turbo is released. Within a single quarter, the API price increase is 83%.

On the application side, since Openclaw has gone viral in China but the official has raised questions about data security, Zhipu took the opportunity to launch a domestic alternative to OpenClaw—AutoClaw, with simple installation, one-click deployment, supplemented by monthly packs with 39/35M tokens, as well as 99 yuan/100 million tokens.

With the explosive growth of Agents and the increase in AI penetration in the IT sector, the company’s stock price rose by 3.5x since February. The core behind this is the change in the pricing logic business model:

With support from a “top-tier model,” the company has already started shifting from a project-based on-prem deployment approach to a cloud-based API interface model. And even after the price increases, the company still says it lacks sufficient compute, reflecting the reality that demand is intense and supplies are tight.

2) Does the quarter’s revenue-generation capacity cover the previous model training investment?

Because the foundation model is updated annually, the model produced using one year of training investment actually only has a one-year service life. Under these circumstances, the model’s economics can be assessed, in part, by the model’s direct and indirect revenue generation in that same year—comparing it with the prior year’s model training investment amount.

For Zhipu, model training spending and R&D personnel spending are mainly captured under R&D expenses (roughly 70% share). Haitun Jun directly uses R&D expenses to see the coverage ability of revenue against its model investment.

Zhipu’s 2024 R&D expenditure is $2.2 billion, while its 2025 revenue is $720 million. It only recovers 33% of the 2024 R&D. For 2025, R&D is $3.2 billion, and if 2026 revenue doubles to around $1.4 billion, you can see the revenue-to-2025 R&D recovery rate rising to 45%, which is basically consistent with MiniMax.

As disclosed in the company’s phone call briefing, its cloud API business’s ARR (annualized monthly revenue) in March has already reached $250 million (RMB 1.75 billion), which is even better than what Haitun Jun expected.

This is very important. Considering that in 2025 the company actually spent RMB 3.2 billion, which means that even before “urgent demand” is fully released in a context of tight supply and inability to recognize revenue, the annualized revenue is already enough to cover 55% of the 2025 R&D expenses. The model is at least on a steady upward business trajectory.

As a comparison, among the Chinese listed company peers that report ARR, Lingyi, which focuses on high-priced video models, had January ARR over $300 million. But because of competition, its full-year guidance is basically just the annualized revenue from January.

MiniMax’s February ARR is $150 million (likely including some continuing revenue that is not pure API). Based on what we see now, both the ARR growth slope and absolute value give Zhipu a clear advantage. Especially since Zhipu’s ARR is accelerating, paired with an upward shift in pricing power and a compute-Tokens supply that remains tight and in high demand—demand has not been fully released yet.

II. Gross margin still in a pain period?

Compared with MiniMax, which monetizes with two legs—both B-side and C-side—Zhipu focuses on B-side business almost exclusively, and most on-prem deployment customers are government agencies and state-owned enterprises.

After DeepSeek, it has become extremely difficult to charge for large models directly. The focus shifts to local adaptation and optimization during the on-prem deployment process. But this deployment approach is labor-intensive (the company’s total headcount increased from 883 around 1H25 to nearly 1,100). In the second half, the disadvantages of resource and headcount investment become even more obvious.

In the second half, the company’s gross profit is only $200 million, up just 30% year over year. Of that, for on-prem deployment business, while revenue grows 57%, gross profit grows only 5%.

And the reason is what Haitun Jun said above: on-prem deployment requires heavy resource input, and scaling up revenue doesn’t come with strong scale effects. As a result, on-prem deployment gross margin directly drops from nearly 60% to 44%. Gross margin continues to decline steadily as revenue expands.

In contrast, although the model’s API interface business starts from a very low base due to fierce competition, it benefits from better scale effects. In the second half, revenue grows 430%, directly lifting gross margin from roughly the zero level to 22%.

So the outcome is: slower growth in high-gross-margin on-prem deployment plus declining gross margin; while low-gross-margin business surges and improves gross margin. Combined, the company’s gross margin in the second half drops to a new low since listing—only 38%, far below market expectations.

III. Revenue generation of $700 million, but losing $3.2 billion! When it comes to dreams, do losses stand aside?

A gross margin of 38% seems not too stressful, because it still hasn’t included the largest investment in large-model startup business models—training expenses—which are included within R&D expenses.

Normally, R&D expenses are 3–5x revenue. So as long as the company is still rapidly using training iterations to improve models, turning profitable is almost impossible (click here to see the reason).

In the second half, Zhipu’s R&D expenses (mainly training costs) are close to RMB 1.6 billion, while revenue for the same period is only $530 million. R&D expenses are three times revenue.

Against such massive R&D spending, changes in other expenses are “minor compared to major.”

Company administrative expenses are $320 million, up 290% year over year, which is relatively faster. But like MiniMax, Zhipu’s selling expense is down 25% year over year. In the second half, it’s only $180 million (also because customer acquisition relies on the model’s own capability rather than on marketing).

Ultimately, the company’s 2H25 gross profit minus the three expense categories results in an operating loss of $1.9 billion. The loss ratio is 354%. After excluding stock-option incentives, it is $1.5 billion RMB—slightly narrowing compared with the over $1.7 billion in the first half.

The company’s full-year adjusted net loss is RMB 3.2 billion, with a loss ratio of 439% (loss is more than 4x revenue), but this is already in a state of rapid narrowing.

The extent of the narrowing loss ratio is even more obvious in the second half: the second half’s adjusted net loss is RMB 1.4 billion. Compared with revenue of RMB 500 million, the loss ratio is 268%.

Haitun Research’s overall view: AI is hot—Is Zhipu even hotter?

Both being the only two listed domestic AI model champions, Zhipu initially had slightly weaker word-of-mouth versus MiniMax. But after just one quarter, Zhipu’s actual moves look more impressive.

In Haitun Jun’s view, the core difference between the two is that Zhipu’s model ranks higher on the Intelligentization Index. In other words, in the B-side productivity domain, intelligent “scarcity” is the core asset, and it is also the key to Token sales pricing power.

Capital’s valuation of the model still places more emphasis on how intellectually scarce the model is—and on, on top of that level, Token sales volume and sales revenue.

Because the company is already publicly listed, it doesn’t need to worry about financing. And given that API prices are rising so sharply right now, while Tokens are still in short supply, financing is even less of a problem.

The company’s big jump in share price after the holidays also shows that the market’s pricing of it has shifted from a discount valuation typical of on-prem deployment AI providers to a valuation model for overseas comparable B-side business models such as Claude.

After the stock price surges 6x in one quarter, the key question is: as of now, the ARR and the ARR growth slope. The company obviously understands what the capital market wants—during the phone call, it was already mentioned that the March ARR for a single type of business, API interfaces, has risen to $250 million, and compute supply is tight and sold out easily.

This kind of guidance makes it very easy for the market to compare Zhipu with overseas peers using their growth and valuation curves: after Anthropic’s intelligence for its model passed the so-called G-point, revenue grew from $100 million to $1 billion in just one year; from $1 billion to $10 billion it also took just one more year.

So the question is: if Zhipu follows a similar revenue growth curve, reaching $1 billion in one year, is there still room for valuation?

As a reference, when Anthropic’s annualized revenue is $1.4 billion, the valuation in the primary market for its financing round is $61.5 billion.

And with the company’s market cap currently at $40 billion, with the next push toward $60 billion, it still can be directly quantified that the model’s intelligentization degree and, more than anything else, the ARR growth slope driven by being well-received carry more weight than Mount Tai.

At least as of now, Zhipu’s momentum seems even more aggressive than MiniMax’s.

Articles:

《Large model losses of 360%, yet MiniMax is still “a hot commodity”?》

《Behind the疯狂高估值: Is MiniMax bubble, or did it touch the future?》

《Deep dive into MiniMax vs. Zhipu: the battle for large-model, compute intensity, and financing staying power?》

**Risk disclosure and disclaimer for this article:**Haitun Research disclaimer and general disclosure