Figured out what "quick money" Zhipu is, and the stock price surged nearly 100 billion Hong Kong dollars in half a day.

Why Does AI Zhipu’s Losses Expand While Its Stock Price Soars?

Image source: Visual China

Text | Yaxuan

Editor | Ye Jinyan

Produced by | Deep Web · Tencent News Xiaoman Studio

“A good business but losing money—that’s when this business grows the fastest.” An investor once explained why some listed companies are favored by investors despite losses.

The latest financial report from Zhipu AI, dubbed the “world’s top large model stock,” vividly illustrates the real-world logic behind this statement.

On the evening of March 31, Zhipu released its first financial report since going public. For the full year 2025, revenue was 724.3 million yuan, up 131.9%; net loss was 4.72B yuan, expanding 59.5% year-over-year; after excluding non-operating losses such as share-based payments and fair value changes of financial instruments, adjusted net loss was 3.18B yuan, up 29.1% YoY.

Normally, continuous loss expansion would trigger a sharp drop in stock prices, but from the stock’s performance on the first trading day after the report, Zhipu’s stock price actually rose.

By the close of Hong Kong stocks on the morning of April 1, Zhipu’s stock price had increased by 31.87%, to HKD 914.5 per share, with a total market value of HKD 407.7 billion; compared to HKD 309.2 billion at the close on March 31, the half-day market cap increased by nearly HKD 100 billion.

Why does Zhipu’s stock “rise even as losses grow”? Combining the financial report and CEO Zhang Peng’s comments on API growth and the shift in the AI paradigm, it becomes clear that Zhipu is not unprofitable but “strategically willing to incur losses.”

Understanding the “Fast Money” Strategy

In 2025, Zhipu’s gross profit was 297 million yuan, up 68.7%, but its overall gross margin dropped from 56.3% in 2024 to 41.0% in 2025. The report explains that this is mainly due to increased revenue share from cloud deployment services and a temporary decline in gross margin for localized deployment services.

Let’s briefly clarify Zhipu’s two core business models:

Cloud deployment services, i.e., standardized API (MaaS) services. The core is packaging the intelligence capabilities of large models (such as GLM-5) into standard interfaces, allowing developers or enterprises to call on demand and pay based on token consumption. This “pay-as-you-go” model has low marginal costs and strong scale effects, validated by leading global AI companies like OpenAI and Anthropic as a core profit driver.

Localized deployment services, which provide private and customized solutions for large clients. This model emphasizes delivery and service, effectively meeting the rigid needs for data security and customization in government, enterprise, and financial sectors, but faces scalability challenges.

In 2025, Zhipu’s revenue from localized deployment was 534 million yuan, accounting for 73.7% of total revenue, down from 84.5% in 2024; cloud deployment revenue was 190 million yuan, accounting for 26.3%, up from 15.5% in 2024.

From the revenue composition, localized deployment remains the main revenue source, but the growth rate of cloud deployment is significantly higher—292.6% in 2025 versus 102.3% for localized deployment.

However, localized deployment requires dedicated teams for customization per client, making marginal cost reductions difficult. This is reflected in gross margins: localized deployment gross margin fell from 66.0% in 2024 to 48.8% in 2025.

In contrast, cloud deployment gross margin improved markedly, from 3.3% in 2024 to 18.9% in 2025, driven by efficiency gains from model inference improvements, increased call volume, and stronger pricing power.

Notably, this margin improvement was not due to price cuts but was achieved through an increase in core API prices amid market conditions.

The report shows that by March 2026, platform users exceeded 4 million, and despite API call prices rising by 83% since the end of last year, market demand still outstrips supply, creating a “compute power panic.”

In response, Zhang Peng explained at the earnings conference, “APIs convert AI infrastructure capabilities into resources for economic operation, not just one-time gains. AI capabilities have shifted from being available and fun to truly solving increasingly complex and important problems, turning token API calls and consumption into real economic value.”

“The More Losses, the More Attractive”

Zhipu’s 292.6% YoY growth in cloud deployment in 2025 further confirms that advanced intelligence is a scarce resource today; whoever controls the upper limit, controls the pricing power.

The financial report states, “If the upper limit of intelligence determines the pricing power of technology, then the scale of token consumption determines the size of commercial value.”

Based on this understanding, Zhipu explicitly presents its business logic toward AGI in the report: AGI commercial value = intelligence upper limit × token consumption scale.

To maintain high-level intelligence in models like GLM-5, Zhipu must sustain continuous generational leadership and technological barriers, which requires heavy investment in R&D, massive data and computing power iterations, and increasing AI talent density to keep elevating the intelligence ceiling.

This precisely explains why Zhipu, despite losing 4.72B yuan in 2025, remains highly favored by capital markets.

The report clearly states that the net loss increase in 2025 was mainly due to sustained R&D investments.

In 2025, R&D expenses were 3.18 billion yuan, up 44.9%, far below the 131.9% revenue growth. This indicates that while maintaining intense R&D to solidify the “intelligence upper limit,” the commercialization of core business has begun to show scale effects.

The report attributes the growth in R&D costs mainly to two factors:

First, increased employee costs, including expanding R&D teams and share-based payments. By the end of 2025, Zhipu had 1,094 full-time employees, with total compensation (including share payments) of 1.36B yuan.

Second, increased costs paid to third-party computing service providers for model iteration and training infrastructure. In Zhipu’s accounting, costs for GPU-based compute services used during model training are included in R&D expenses; long-term GPU leasing costs are capitalized.

In other words, Zhipu’s R&D spending mainly invests in AI talent and computing resources—necessary costs to maintain high-level model intelligence and control “intelligent pricing power.”

To focus funds more directly on R&D and business growth, Zhipu adjusted its compute procurement model in 2025, favoring service procurement combined with some equipment leasing.

The report shows that Zhipu’s capital expenditure in 2025 was about 74.7 million yuan, a reduction of approximately 83.8% from 462.3 million yuan in 2024. This significant cut in capital spending further improves the efficiency of core R&D investments.

Additionally, a key logic appreciated by the market is that Zhipu has reduced dependence on any single major client. The report indicates that in 2024, the largest client accounted for 19% of revenue, while in 2025, no client contributed more than 10%.