Zhipu is rushing forward while losing blood.

Ask AI · Zhipu’s R&D investment far exceeds revenue—can this model last?

Produced by | Huxiu Technology Team

By | Song Sihang

Edited by | Miao Zhengqing

Cover image | Visual China

If there were a business where “for every 1 yuan of revenue earned, 4.4 yuan is invested in R&D,” would that be a good business?

Zhipu AI (below, “Zhipu”), the first publicly listed global large-model company, is trying to answer this question. On March 31, 2026, Zhipu released its full-year 2025 results, which is also its first set of financial statements since listing.

The report shows that in full-year 2025, Zhipu generated total revenue of RMB 724 million, up 131.9% year over year; selling costs rose 213.3% to RMB 427 million. With such revenue and cost performance, Zhipu’s gross profit grew 68.7% year over year to RMB 297 million, and its full-year consolidated gross margin also reached 41%. Meanwhile, for MiniMax, another listed large-model company, Zhipu’s gross margin is already far higher than MiniMax’s—25.4%.

If you break down the revenue composition, the 41% gross margin behind it reflects a divergence trend: the gross margin for its localized deployment business fell from 66.0% in 2024 to 48.8% in 2025, while the gross margin for its cloud deployment business rose from 3.3% in 2024 to 18.9% in 2025. (Huxiu note: localized deployment refers to deploying large models locally; cloud deployment refers to open platforms and APIs.) Judging from this, Zhipu has already sufficiently proven its own room for profitability.

But it is still operating at a loss.

Affected by RMB 3.18 billion in R&D expenses in the same period, Zhipu’s adjusted net loss reached RMB 3.182 billion, with the loss level up 29.1% year over year. The loss in the same period was equivalent to 4.39 times Zhipu’s total revenue, and even 10.7 times its gross profit. It is worth noting that the losses mainly come from R&D spending. According to the financial report, Zhipu’s R&D costs in 2025 were RMB 3.184 billion, up 44.9% year over year; while capital expenditures in 2025 were RMB 74.70 million.

In the financial report, it explained that the growth in R&D costs mainly came from:

(1) Increased employee costs, including expanding the R&D team and higher share-based payment expenses;

(2) Computation service fees paid to third-party computing power providers, including costs related to iterating models and investing in more advanced model-training infrastructure.

But it is worth mentioning that the computation costs used for large-model training are not included in R&D expenses; instead, they are recorded separately as capital expenditures in the form of compute-power leasing. In Zhipu’s context, the former—compute costs for large-model training—refers to the cost of calling GPU resources from the computing supplier based on the duration of model training; this flexible spending is booked into R&D costs. In contrast, when GPU resources are locked in and long-term contracts are signed with a specific supplier, it is treated as capital expenditures.

Compared with MiniMax, Zhipu is on a larger overall scale. This is mainly due to differences in business composition and organizational structure between the two companies. For example, Zhipu’s headcount is twice that of MiniMax, which also leads to higher R&D spending and more severe losses; while the latter has higher efficiency per employee.

One notable point in this financial report is that, like MiniMax, Zhipu has also tasted the “lobster dividend.”

Starting from Q1 2026, Zhipu’s performance growth has relied mainly on AutoClaw it launched in March—one-click deployment of lobsters.

According to Zhipu CEO Zhang Peng, in the first quarter, Zhipu’s API pricing increased by 83%. But it also happened to coincide with the timing of a surge in demand. At that time, the popularity of lobsters had already lasted for a month. In the half month after the price increase, Zhipu started deploying lobsters. So even though prices rose, the call volume for Zhipu’s GLM models still grew 400%. According to the financial report, two days after the plan launched, subscription users exceeded 100,000; after 20 days, subscription users surpassed 400,000.

Corresponding to this is the indicator of Zhipu’s profitability—namely the MaaS platform, where Zhipu is putting its weight behind it. It is said that the MaaS API platform achieved ARR of RMB 1.7 billion (about US$250 million), up 60-fold year over year.

That is to say, this financial report from Zhipu essentially proves its room for profitability on one hand, while showing that losses have not stopped on the other.

Zhipu’s growth logic has changed, but it hasn’t been fully rebuilt

From the overall income structure, the key variable in this financial report is not actually total revenue itself, but the source of revenue. By dissecting subtle changes in the revenue sources, you can glimpse Zhipu’s new growth logic and its sustainability.

Looking more closely, Zhipu’s growth focus has started to tilt toward the cloud—i.e., MaaS. This segment accounts for 26.3% of business share; in 2024, cloud deployment accounted for only 15.5% of total revenue. After the release of this financial report, Zhipu also claimed that the company’s strategic focus will continue to be on MaaS.

However, even though the share of cloud deployment has increased significantly in the numbers, several variables are especially critical.

First, the most core driver comes from APIs. In other words, the growth this round for Zhipu is fundamentally growth in call volume.

Among that, lobster (OpenClaw) is the most direct variable. With agents starting to automatically execute tasks, a single demand often corresponds to multiple rounds of calls, which multiplies token consumption, and in turn drives up API call volume.

Second is the main revenue source for MaaS. The financial report states that among ten internet companies, nine have already connected Zhipu’s model.

There is a change worth noting here: these internet companies basically all have their own large models, but they do not rely entirely on their own models; instead, they call different models for different businesses. That means that even if they have developed large models in the short term, they will still choose Zhipu in specific scenarios. Of course, this does not mean that these nine internet giants will choose this strategy long term.

And the call volumes from these companies basically make up about half of Zhipu’s MaaS revenue. That is to say, once Zhipu loses any one of these customers, it would deal a serious blow to its current MaaS business.

Third, MaaS growth also comes from exporting tokens overseas. Over the past year, Zhipu has carried out cooperation with multiple Middle Eastern and Southeast Asian countries, exporting model capabilities to local markets. Fundamentally, this is also revenue achieved in the form of token calls.

Overall, a clear signal released by this financial report is that Zhipu is switching the narrative of growth—from heavy localized deployment to selling models, i.e., selling tokens.

But in terms of outcomes, although Zhipu’s current main revenue still relies on localized deployment, the MaaS model has shown a trend of sustainable growth.

On this basis, Zhipu also proposed a new concept: TAC (Token Architecture Capability, token architecture capability).

According to its definition, TAC consists of three parts: intelligent call volume, intelligent quality, and economic conversion efficiency. Simply put, it means how many tokens are called, whether those calls are effective, and whether they ultimately can be converted into revenue.

In the author’s view, after the “lobster” event, the industry gradually reached a consensus on tokens: once large models have long-horizon task execution capability, calls are no longer just one-time input-output; they are organized into a continuously running system.

So, behind a single task, it often corresponds to multiple rounds of calls, tool calls, and even self-checks. Tokens are no longer merely consumed—they are “orchestrated,” meaning how the user organizes calls to the large model.

And why TAC is proposed at this point is not hard to understand either.

Over the past two years, competition in the large-model industry mainly revolved around parameter scale, model capability, and price; but as the price war is drawing to a close and model capability gradually converges, agent application scenarios have started to surge, making these metrics increasingly unable to explain differences in companies’ growth.

Against this backdrop, Zhipu needs a new set of metrics to answer a more realistic question: when model capabilities differ only slightly, where does growth come from?

Zhipu’s “cost trap”

If you zoom out from Zhipu to the entire industry, you will find that the business model of large models has begun to converge.

Apart from Step-Function Star, the other three foundation-model companies all have core revenue converging toward API calls.

Whether it is Zhipu, MiniMax, or the dark side of the moon, all are moving toward using MaaS to capture growth. But at least for Zhipu, this path did not exist from the beginning.

Take Zhipu as an example. In its early business, ToG and private deployment accounted for a very high share, and the projects had clear characteristics. Only about half a year before listing, in order to make the business model more sustainable and provide room for scale imagination, Zhipu began to clearly pivot toward MaaS, shifting its growth focus to cloud API calls.

Based on results, this transformation has indeed brought changes: MaaS share increased, tokens became the core metric, and the revenue structure began to move toward platformization.

However, given Zhipu’s current architecture, a structure where localized deployment exceeds cloud deployment is difficult to change in the short term.

Zhipu’s current MaaS growth also heavily depends on a small number of major customers.

The financial report shows that a substantial portion of Zhipu’s API revenue comes from internet blue-chip companies. These companies may have developed their own models, but in specific business scenarios they choose to call external model capabilities. This “multi-model calling” pattern does provide stable demand for MaaS. The problem is that this is not equivalent to truly scalable growth in the real sense.

On one hand, the top customers contribute most of the call volume; on the other hand, the long-tail market has not truly been opened. In other words, while the platform-like form of MaaS has appeared, platform scale has not been established.

And this also points to another even more core issue: the cost per token and the revenue structure.

The financial report shows that in 2025, Zhipu’s full-year loss was RMB 4.718 billion, up 59.5% year over year; of this, R&D expenses were RMB 3.18 billion, up 44.9%; capital expenditures were RMB 74.70 million, down about 83.8% year over year. The former refers to model training costs and employee costs, while the latter comes from costs such as compute-power leasing. In 2025, Zhipu adjusted its approach to compute-power procurement, changing from what was previously more fixed compute-power leasing into a model combining compute-power leasing and service procurement, which is why capital expenditures fell significantly.

Then, by combining MaaS growth with these two sets of data, you can observe a very direct logic chain:

If a company wants to drive MaaS growth, it must rely on model capabilities; and improving model capabilities depends on continuously increasing R&D investment. But the problem is that R&D and compute costs do not naturally decline as the scale of calls expands.

In other words, the premise of revenue growth is itself pushing up costs.

This is what puts large-model companies into a structural dilemma: to get more calls, they must continuously improve model capabilities; and to improve model capabilities, they must keep increasing investment.

This is why the situation today is that the faster the growth, the more cost pressure there is.

From this perspective, the problem is no longer just one company—Zhipu. It is a common constraint facing the entire large-model industry.

Until this problem is solved, MaaS can bring growth, but it is hard to bring profits.

Why Zhipu wants to benchmark Anthropic

At the earnings release conference call held on the evening of March 31, before Zhipu CEO Zhang Peng reported the results, he specifically mentioned Anthropic, an American AI unicorn. Anthropic’s ARR grew from US$1 billion at the end of 2024 to US$9 billion at the end of 2025.

In fact, almost all leading large-model companies are trying to follow the American path.

For example, the dark side of the moon has focused on OpenAI, pursuing a “model capability + products + subscriptions” route; while Zhipu and MiniMax are trying to move closer to the Anthropic model—emphasizing base model capability, outputting inference compute via APIs, and building a developer ecosystem.

But no matter which path is chosen, fundamentally, it is turning the model into infrastructure and achieving scalable revenue through calls.

This path has already been preliminarily validated in the U.S. If either OpenAI or Anthropic can show that when model capability is strong enough, a developer ecosystem can form a positive feedback loop. The problem is that this path is hard to replicate in China.

First, the pricing system is different.

In the U.S. market, enterprise customers and developers are more willing to pay for capabilities, so model capability can be converted into premium pricing; but in China, prices are quickly driven down from the start. After two years of price wars, tokens only gradually evolved into “basic resources.”

Second, the structure of demand is different.

The large-model ecosystem in the U.S. relies more on long-tail developer demand; while in China, more calls are concentrated in major customers—such as internet giants and government/enterprise customers. Under this structure, MaaS is more like centralized procurement “—not driven by a developer ecosystem.”

Third, there are differences in cost and supply. Compute supply, chip structures, and the overall cost environment make it more difficult for domestic model companies’ costs to decline with scale.

From this perspective, Zhipu’s dilemma becomes easier to understand.

Looking back at the development paths of the internet and cloud computing, profitability at the infrastructure layer is often built after the application layer surges.

Similarly, that also means that at the current stage, whether it is Zhipu or other large-model companies, they need to wait until application scenarios have been continuously validated before scale effects may emerge.