Four years of massive losses totaling 8.5 billion yuan, yet the stock price has surged dramatically. Can Zhipu sustain a market value of 300 billion yuan?

Ask AI · How does the算力缺口 (computing power gap) of Zhipu affect its commercialization prospects?

“The world’s leading large model stock” How to solve the算力缺口?

Author | Yao Yue

Editor | Yu Xing

Source | Yema Finance

As “the world’s leading large model stock,” Zhipu (2513.HK) has just delivered its first full annual report since listing in Hong Kong.

In 2025, the company’s net loss is 4.72B yuan, and combined with losses from 2022 to 2024, the total loss over four years is about 8.5 billion yuan. Yet, this continuous loss report has instead been highly favored in the secondary market, with the stock price soaring over 30% in a day, and market capitalization once surpassing 400 billion Hong Kong dollars, with a price-to-sales ratio even leaving industry leader OpenAI far behind.

Behind the high valuation from the market, Zhipu is transforming towards cloud API and enterprise intelligent agent businesses, and the market is optimistic about its commercialization potential. But whether all this can truly be realized depends largely on whether the算力 (computing power) can keep up and costs can be reduced. Not long ago, when GLM-5 was launched, Zhipu experienced service instability due to insufficient算力, and publicly apologized, which was seen as exposing the issue that its算力 mainly comes from third parties and lacks self-built infrastructure.

As the first batch of restricted shares is about to be unlocked in July, Zhipu will face a stress test in the secondary market. Industry experts believe that if the算力 bottleneck cannot be resolved in time and profits do not improve significantly, the current high valuation may see a correction. Additionally, with the unlocking increasing circulating shares significantly, stock price volatility and subsequent financing pressures could further intensify.

As of the close on April 2, Zhipu’s stock price was 779 HKD per share, down 14.86%, with a total market cap of 347.3 billion HKD (about 305.01B RMB).

This image may be AI-generated

01
2025 Loss of 4.72B yuan

Year-on-year increase of 59.5%

On the evening of March 31, Zhipu released its first full annual report since listing—revenue grew rapidly, but losses expanded significantly.

Zhipu mainly has three business segments—Open Platform and API, its own large model “connected” to clients and charged by usage; enterprise-level intelligent agents, selling AI employees to companies; and enterprise-level general large models, selling AI brains to enterprises.

The financial report shows that, in 2025, Zhipu achieved total revenue of 720 million yuan, a year-on-year increase of 131.9%. Looking at the segmented businesses, revenue from open platform and API was 190 million yuan, up 292.6%; enterprise intelligent agents earned 166 million yuan, up 248.8%; and enterprise general large models brought in 366 million yuan, up 70.5%.

But the other side of high growth is expanding losses and cash “loss of blood.”

The financial report shows that, in 2025, Zhipu’s loss was 4.72B yuan, an increase of 59.5% year-on-year; adjusted net loss was 3.18B yuan, up 29.1%. According to Wind data, from 2022 to 2024, Zhipu lost 143 million, 788 million, and 2.96B yuan respectively. Adding up over four years, the total loss is about 8.5 billion yuan.

Zhipu’s services are realized in two ways—cloud deployment: like electricity, no need to buy generators, just turn on the switch and pay for usage; localized deployment: like buying a generator yourself, placing it in your yard, generating and using power yourself, others cannot use it.

In 2025, Zhipu’s overall gross profit margin sharply declined from 56.3% to 41%. Among them, although cloud deployment gross margin increased from 3.3% to 18.9%, it remains far below traditional localized business; meanwhile, traditional local deployment, due to higher customer requirements, increased cost input, leading to gross margin dropping from 66% to 48.8%. The rapid expansion of low-margin businesses, compared to the slower growth of high-margin ones, means profit improvement has not kept pace with revenue growth.

More importantly, R&D and算力 (computing power) “big consumption” are eating into profits. In 2025, Zhipu’s R&D expenditure reached 3.18 billion yuan, equivalent to 4.4 times its revenue during the same period, meaning that for every yuan earned, 4.4 yuan is spent on R&D. Most of this funding flows into model training,算力 costs, and high-end talent salaries. Large model training and inference are industry “black holes” for算力, with flagship models like GLM-5 increasing single inference capacity, further driving up算力 costs.

To ensure operations, Zhipu borrowed an additional 552 million yuan in 2025 compared to 2024, with total borrowings reaching 690 million yuan by the end of 2025, all with interest. But under the huge daily expenses, Zhipu’s cash reserves did not grow positively, by the end of 2025, Zhipu’s cash and cash equivalents were 2.26B yuan, down 10.1 million yuan year-on-year.

Based on this, Zhipu raised about 4.17B HKD (roughly 3.67B RMB) in its Hong Kong IPO in January, which was clearly very timely.

Yuan Shuai, co-founder of New Zhituo New Quality Productivity Salon, said that although some funds were raised through listing, given the current annual loss rate of nearly 5B yuan, the existing capital reserves are obviously insufficient for long, and urgent financing needs have almost become inevitable.

02
Loss report unexpectedly triggers secondary market rally

However, even this first loss report, which shows expanding losses, still excited the secondary market.

On April 1, the day after the 2025 annual report was released, Zhipu’s stock price surged 31.94%, closing at 915 HKD per share, and market cap once exceeded 400 billion HKD, reaching 407.95B HKD (about 52.5 billion USD; 359.26 billion RMB). With revenue of 724M yuan, its P/S ratio is about 500 times, meaning at its peak, the market was willing to value each 1 yuan of profit at about 500 yuan.

How “exaggerated” is this? Compared to overseas AI giants, OpenAI’s P/S ratio is about 35-65 times; Anthropic’s is about 27 times.

In the eyes of outsiders, Zhipu’s capital narrative has also shifted. Previously, large model companies were seen as “money-burning beasts” that bought technology with cash, now Zhipu’s growth momentum is rapidly shifting from project-based local deployment to more scalable cloud API and intelligent agent businesses. The financial report shows that in 2025, revenue share of Zhipu’s open platform and API increased from 15.5% to 26.3%; enterprise intelligent agents from 15.2% to 22.9%; and enterprise general large models’ revenue share decreased from 68.7% to 50.4%.

According to “Cailian Press,” at Zhipu’s 2025 performance briefing, CEO Zhang Peng said that in Q1 2026, Zhipu’s API call pricing increased by 83%. Even so, the market still shows signs of demand exceeding supply, with call volume increasing by 400%.

2025 saw increasingly fierce industry price wars. Zhang Peng emphasized that this information indirectly indicates that Zhipu’s technology value has been recognized by the market, and it marks a move out of the low-level price war quagmire.

03
How to solve the算力缺口?

However, while investors dream of high growth, a harsh reality is that：算力 (computing power) is a key variable determining the ceiling of this story. Imagine a supercar with an ultra-powerful engine, but unstable fuel supply—destined to struggle to go far. And Zhipu recently apologized for its算力缺口.

On February 12, Zhipu released its flagship model GLM-5. Due to significant performance improvements, user enthusiasm far exceeded expectations, and surging traffic directly challenged Zhipu’s算力—to cope with the higher算力 consumption of GLM-5, the company designed layered consumption strategies of 3x during peak times and 2x during off-peak times. Despite this, server expansion could not keep up, and many paying users could not use the core model normally.

On February 21, Zhipu quickly apologized for the mistake. Despite launching “bleeding” compensation, on the first trading day after the apology, Zhipu’s stock price plummeted 23%, evaporating over 70 billion HKD in half a day.

Moreover, entering 2026, the entire算力 demand is undergoing a dramatic change. The explosive application of AI intelligent agents—OpenClaw—has ignited a狂潮 (craze) of算力 consumption. On March 10, Zhipu announced joining the “Lobster Group,” officially launching AutoClaw (Chinese name: Ao Long), one of the first domestic “one-click install” local versions of OpenClaw.

The 2025 annual report shows that Zhipu’s算力 mainly relies on third-party suppliers, with no large-scale self-built算力 centers. As scale expands,算力 demand grows rapidly, and from 2025, Zhipu will no longer lease equipment in large quantities but mainly purchase算力 services directly. This means the marginal cost does not decrease with scale.

To address the算力缺口, Zhipu is promoting domestic算力 adaptation, such as deep adaptation of models like GLM-5 to domestic chips, entering the co-design stage of hardware and software, and seeking “算力 partners” across the network.

Yuan Shuai said that, unlike overseas giants building their own算力 centers and achieving自主可控 (independent controllability), Zhipu lacks self-built infrastructure, and in terms of算力 cost, supply stability, and technical adaptation, it does not have advantages, making it difficult to compete with giants. This could become a key factor restricting its long-term development.

Additionally, it’s worth noting that Zhipu faces increasingly fierce competition: internet giants are heavily blocking, and startups are closing in, squeezing the market space continuously.

04

The first wave of stock unlocks is imminent

What about the A-share listing process?

As of April, about 21-22% of Zhipu’s total share capital is nominally freely tradable (not restricted shares), but since most are held long-term, the actual freely tradable shares in the market are only about 2.67%.

Based on the April 1 closing price of 915 HKD per share, Zhipu’s total market cap exceeds 400 billion HKD, but the real tradable market cap is only about 10.7B HKD, a very small free float, and combined with the frenzy in the AI sector, even slight capital inflows can trigger stock price surges. However, this situation will be broken when the first restricted shares are unlocked in July, facing a severe test of whether original shareholders will sell collectively.

According to Wind data, 25.6816 million shares, accounting for 5.76% of total shares, will be unlocked by July 8. At 915 HKD, the unlock market value is about 23.5 billion HKD, roughly 2.2 times the current free float market cap. Additionally, 178 million shares, accounting for 39.99%, will be unlocked by January 8, 2027.

Image source: Wind Financial Terminal

Bai Wenxi, Vice Chairman of the China Enterprise Capital Alliance, believes that these investors have accumulated high unrealized gains and have strong expectations for cashing out. If, at that time, Zhipu’s API growth slows, gross margin does not improve significantly, and losses continue to expand, investor confidence could weaken, and the liquidity surge after unlocking could cause large stock price fluctuations. With a small free float, the impact of unlocking will be amplified.

Bai Wenxi said that, continued losses combined with large-scale unlocking and selling pressure will significantly increase subsequent financing costs, and valuation of equity financing may be pressured, with stricter debt financing terms. The company needs to communicate with cornerstone investors in advance to lock in commitments or consider phased unlocking arrangements to smooth the impact.

On the other hand, Zhipu is pushing for A+H dual listing. It is already listed in Hong Kong, and according to public information, A-shares on the STAR Market are in IPO counseling, but have not yet submitted formal application materials. If it successfully returns to A-shares, it will open a new financing channel.

What AI products do you usually use? What’s your impression of Zhipu? Feel free to leave comments.

Author’s statement: Personal opinions are for reference only.