Một góc nhìn từ người nước ngoài về Zhipu AI: Mô hình miễn phí, thua lỗ khủng, vì sao vốn hóa từng vượt Meituan?

Author: Robonaissance

Compiled by: TechFlow

Introduction: Zhipu AI's Hong Kong-listed shares surged 25 times in the first six months after its IPO, briefly surpassing Meituan's market cap. Yet in 2025, its revenue was RMB 724 million, with a loss of RMB 4.72 billion. Its flagship model, GLM-5.2, is available for free download under the MIT open-source license. The market isn't crazy; it's pricing rarity, sovereignty, and a tiny free float. A Tsinghua pedigree, state-backed capital, and 73.7% of revenue from state-owned enterprise private deployments—this is the real product Zhipu is selling.

On July 2, 2026, shares of the world's first publicly listed AI lab plummeted nearly 17% in a single day. Six days later, on July 8, when the lockup period expired, freeing approximately HK$46 billion in frozen shares, the stock surged 13%. Within 24 hours, the company seized the rally to issue $4 billion in new shares.

Zhipu AI, listed in Hong Kong as "Knowledge Graph Technology," was rated by Bloomberg as Asia's most volatile stock. But volatility isn't a side effect; the volatility itself is the mechanism.

The underlying business is more surreal than the candlestick chart. In 2025, Zhipu's revenue was RMB 724 million (~$105 million). Its loss was RMB 4.72 billion (~$650 million). R&D spending was RMB 3.18 billion, 4.4 times its annual revenue. The flagship model GLM-5.2 is under the MIT open-source license; anyone can download the weights, run inference locally, fine-tune it, and build commercial products without paying Zhipu a cent.

At the end of June, the market valued the company at HK$1 trillion (~$128 billion). That's higher than Meituan—a company delivering food to hundreds of millions of people that actually turns a profit.

A simplistic reading is that the market is insane. A more useful reading is that the market is pricing something real, but that something isn't on the P&L statement. It's pricing scarcity, sovereignty, and a free float small enough to be pushed around. This is a story about these three things, and what happens when the world finally puts a public price on a frontier AI lab, only to find the price reflects almost nothing about the lab itself.

The Tsinghua Pedigree

Zhipu didn't start as a startup; it began as a university research group. This distinction explains most of the company's subsequent form.

The Knowledge Engineering Group at Tsinghua University, internationally known as THUDM, had been researching knowledge graphs and language models long before they became popular. In 2019, professors Tang Jie and Li Juanzi spun out the work to form the company. The architecture they brought was called GLM, a General Language Model, which became both the company's technical identity and the origin of its name.

This background brought two things, though only one is often written about.

First, there's the technology. In March 2023, when most Chinese AI companies hadn't yet released anything developers could use, Zhipu released ChatGLM-6B, an open-source conversational model small enough to run inference on a single consumer-grade GPU. It became one of the highest-downloaded models that year and the first widely available Chinese instruction-tuned large language model. Enthusiasts fine-tuned it on their laptops, university labs used it for courses, and companies dissected it to understand its principles. The habit of giving away models for free has been there from the start, and the reason is entirely unromantic: free release is how an academic spin-out gets noticed.

The second thing is trust, which later became the business itself. Zhipu became one of the "Six Little Dragons"—the cohort of Chinese large model startups that emerged in the generative AI wave. Before its IPO, it assembled an unusually broad investor list: Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Hillhouse Capital, Qiming Venture Partners, various Chinese local government funds, and Saudi Aramco's Prosperity7 Ventures, totaling about $1.5 billion. A Tsinghua spin-out with state-backed investors on its cap table means Chinese state-owned banks can make procurement decisions without anyone needing to justify the choice. This access isn't a soft advantage. Look at the revenue structure, and you see it is the entire business engine.

Zhipu has fewer than 900 employees, about three-quarters of whom are researchers. The CEO is Zhang Peng, Tang Jie is the core scientist, and the Chairman is Liu Debing. For a company once valued higher than Meituan, this is a very small building, filled with academics.

What Exactly Does Zhipu Sell?

In 2025, of Zhipu's RMB 724 million in revenue, RMB 534 million came from private deployment, accounting for 73.7%.

This single number reconstructs the entire company.

Private deployment means Zhipu's engineers go to a client's facility, install the GLM model suite onto the client's own servers and intranet, ensuring data never leaves the premises. They fine-tune the model with the client's data, integrate it with existing systems, and stay until it works. Then they move on to the next client and do it all over again.

The clients are Chinese state-owned enterprises, banks, and government agencies. These organizations cannot put sensitive data on someone else's cloud, and they won't buy foreign models at any price. For them, a model with a Tsinghua pedigree, state-backed investment, and domestic deployment isn't one option among several; it's the only option. This is what the Tsinghua lineage buys, converted into invoices.

The remaining RMB 190 million, or 26.3%, came from the cloud business: APIs, developer platforms, the parts that run like software. This segment is growing rapidly, its revenue share rising from 15.5% in 2024. Gross margin climbed from 3.3% to 18.9% as inference optimization and scale pushed down per-token marginal cost.

But the company's shape is determined by that 73.7%, and that shape has a problem. Overall gross margin fell from 56.3% in 2024 to 41.0% in 2025. The gross margin for the private deployment segment itself dropped from 66.0% to 48.8%. As the business grows, margins contract because growth means hiring more engineers to sit in more buildings. There's no version of private deployment where serving the tenth client is cheaper than serving the first. Labor has no cost curve, only headcount.

There's no question about growth. Zhipu's revenue was RMB 57.4 million in 2022, RMB 124.5 million in 2023, RMB 312.4 million in 2024, and RMB 724 million in 2025. This is a company roughly doubling every year. By revenue, Zhipu is China's largest independent large model developer. The trajectory is exactly what a believer would want to see.

The problem lies in the bill below, which dictates everything. Zhipu's computing costs paid to third parties were RMB 14.6 million in 2022, RMB 311.7 million in 2023, and RMB 1.55 billion in 2024. In just the first half of 2025, according to the prospectus figures, it was RMB 1.15 billion.

Look at these two lines together. Between 2022 and 2024, Zhipu's revenue grew about 5 times. Over those same three years, the computing bill grew well over a hundredfold. And in the first six months of 2025, it spent more on computing alone than it earned in all of 2024.

These two lines are not converging.

Traditional software is written once and copied at zero marginal cost, making software companies the most profitable businesses in history. Large models break this. You write once, but you pay again every time someone uses it. Revenue grows linearly; computing consumption grows along a curve, and the curve steepens as context windows lengthen and reasoning chains grow. In 2025, for every dollar Zhipu earned, it paid far more than one dollar to chip vendors and cloud providers.

The Open-Source Paradox

Zhipu's strongest model is free.

GLM-5.2, released in mid-June 2026, supports a context window of up to 1 million tokens and is licensed under MIT. Download the weights, run on your own hardware, modify, build a product, sell it—never pay Zhipu a cent. This isn't a watered-down community edition; this is the flagship, the one the company uses to benchmark against American frontier models.

The obvious question is, how can this possibly be a business?

The answer is that open source is a distribution strategy Zhipu can afford precisely because its revenue doesn't come from selling access to the model. Revenue comes from selling deployment, integration, and services. Open-source weights are marketing; private deployment contracts are the product. Giving away the model for free doesn't cannibalize revenue, because the paying customers were never going to rent the model via API anyway. They were always going to pay someone to come to their building and install it.

What free release buys is coverage. According to the company, over 4 million registered corporate and developer users across 218 countries and regions, integrating with nine of China's top ten internet companies. It buys developer mindshare, which is the raw material for API revenue. And it buys a specific kind of credibility: a model whose weights have been scrutinized by the entire world is one a bank's risk committee can approve without taking anyone's word for it.

The proof the strategy works isn't downloads—downloads are easily gamed and hard to monetize. The proof is price. While Chinese competitors are slashing prices to grab market share, Zhipu raised its API prices by 83%, and demand still outstripped supply. The open platform's ARR reached RMB 1.7 billion (~$240 million), up 60x year-over-year, a figure Zhang Peng gave on the company's first earnings call as a public company.

A company that can raise prices in a price war has something the price war cannot touch.

What changes the demand curve is agents. The GLM-5 series is optimized for long-horizon software engineering; the model can work continuously for hundreds of iterations instead of stopping after answering one question. Zhipu's programming packages integrate into tools developers already use. When a code agent runs autonomously for an hour, it doesn't consume the token volume of a single query; it consumes thousands of times that amount. Chairman Liu Debing's argument is that the resulting usage and price increases are persistent, not spiky, driven by models becoming more capable and users letting them do more work.

That's the bull case in one sentence: agents are token furnaces, and Zhipu sells tokens.

The Float Machine

But none of this explains how a stock went from HK$116.20 in January to an intraday high of HK$2980 on June 22, a 25x increase in under six months. To explain that, you have to look at the plumbing.

Zhipu listed on January 8, 2026, at an issue price of HK$116.20, selling about 43 million shares including the over-allotment option, roughly 9.65% of its share capital. Eleven cornerstone investors took about HK$2.98 billion, close to 70% of the offered shares. Cornerstone investors are large institutions the Hong Kong issuer brings in before the IPO: they commit to buying a large block, guaranteeing their allocation, in exchange for agreeing not to sell for six months. Retail investors oversubscribed the remaining portion by over a thousand times.

Do the math. On the first day, the truly tradable shares numbered about 17.35 million. That's less than 4% of the company.

A stock with a 4% float doesn't have a normal stock's price. It has a clearance level between a very small number of people willing to sell and whatever demand materializes. And the demand that materialized in the first half of 2026 was every investor on the planet who wanted Chinese frontier AI exposure and, before January, had no publicly listed pure play. Not DeepSeek, that's private. Not Moonshot AI, that's private. Not Huawei, which doesn't list and doesn't sell models. Just Zhipu, and MiniMax listing a day later—that was the entire menu.

UBS was blunt: the valuation reflects a scarcity premium and a limited number of tradable shares. Bloomberg later observed that Zhipu's stock was Asia's most volatile, largely due to the tiny float.

This machine ran for six months. Then July arrived and ran it again, in public view.

On July 2, as the cornerstone lockup approached, the stock fell nearly 17% in a single day. Nothing happened except the anticipation of supply. It closed at HK$1,754, market cap falling below HK$800 billion. The trade was too crowded, the float too thin; the mere prospect of 25.68 million shares—5.8% of the company—coming unlocked wiped out about a sixth of the market cap in one day.

Then on July 7, the lockup expired, and cornerstone investors didn't sell. Nearly 70% of them committed to holding. On July 8, the stock rose 13.35%, adding over HK$100 billion in market cap in a single day, because not selling was interpreted as a vote of confidence.

Within 24 hours, Zhipu issued new shares into the rally. It placed approximately 19.8 million new shares at HK$1,588 per share, raising about HK$31.4 billion (~$4 billion), one of the largest placements in Hong Kong this year, and over six times the size of Zhipu's own IPO. CICC and CCB International were bookrunners. The price was set at a ~13% discount to the previous close, the discount needed for institutions to take the stock at that level.

After the lockup release and after the placement, only about 13.5% of Zhipu's outstanding shares were freely tradable.

The company didn't fix the float; it monetized the float.

The controlled experiment ran the next day. MiniMax, another Chinese model developer listed in Hong Kong in January, had its own lockup expiration on July 9. The founder extended the lockup; strategic shareholders pledged not to sell. The stock still fell over 20% intraday.

The difference isn't in the underlying architecture—the architecture is largely the same. The difference is that MiniMax tried to raise prices for its M3 model, the market rejected it, and it had to cut them. Zhipu raised prices 83%, and the market accepted it. A thin float amplifies whatever market belief exists; it doesn't manufacture belief out of nothing. When belief is there, a thin float turns a good quarter into a 25x gain. When belief is absent, the same thin float turns a lockup expiration into a rout.

Training Without Nvidia

For anyone tracking the Chinese AI tech stack, the most important claim is one Zhipu made quietly.

Reportedly, the open-source flagship model GLM-5, released in February 2026, was trained and deployed on Chinese accelerators rather than Nvidia hardware: Huawei's Ascend, and chips from Cambricon, Moore Threads, and Kunlunxin. On the earnings call, Zhang Peng stated that since February, Zhipu has been accelerating the use of domestic chips to meet surging computing demand. The R&D budget includes co-design work for domestic chip adaptation. The company built its own asynchronous reinforcement learning framework, Slime, partly to ensure the training pipeline could run on the hardware it was actually allowed to buy.

This path wasn't so much chosen as it was chosen for Zhipu. In January 2025, the US Commerce Department added Beijing Zhipu Huazhang Technology and its subsidiaries to the Entity List, citing concerns the company was helping advance China's military modernization through AI. Zhipu disputed the justification and stated it does not rely on American large model technology. Regardless of how one views the designation, its practical