A foreigner’s take on Zhipu AI: Free models, massive losses—why was its market cap once higher than Meituan?

Original Author: Robonaissance

Original Compilation: Deep Tide TechFlow

Introduction: Zhipu AI's Hong Kong-listed shares surged 25 times in the first half of the year, at one point surpassing Meituan's market cap. Yet its 2025 revenue was just 724 million RMB, with losses of 4.72 billion RMB. Its flagship model, GLM-5.2, is available for free download under the MIT open-source license. This isn't the market going crazy; it's pricing scarcity, sovereignty, and a tiny, floatable circulating supply. With its Tsinghua pedigree, state-backing, and 73.7% of revenue from privatization deployments for state-owned enterprises—this is what Zhipu is actually selling.

On July 2, 2026, shares of the world's first publicly listed AI lab plummeted nearly 17% in a single day. But on July 8, six days later, as the lock-up period expired and approximately 46 billion HKD in frozen shares were unlocked, the stock surged 13% instead. Within 24 hours, the company took advantage of the rally to issue 4 billion USD in new shares.

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

The underlying business is more surreal than the candlestick charts. In 2025, Zhipu's revenue was 724 million RMB, about 105 million USD. It posted a loss of 4.72 billion RMB, approximately 650 million USD. R&D spending hit 3.18 billion RMB, 4.4 times its annual revenue. Its flagship model, GLM-5.2, uses the MIT open-source license. Anyone can download the weights, run inference themselves, fine-tune it, and build commercial products without paying Zhipu a cent.

By late June, the market valued the company at 1 trillion HKD, roughly 128 billion USD—higher than Meituan, which delivers food to hundreds of millions of people and actually turns a profit.

The simple interpretation is that the market has gone mad. A more useful interpretation is that the market is pricing something real, but that something isn't on the income statement. It's pricing scarcity, sovereignty, and a float so small it can be pushed around. This is the story of those three things, and of what happened when the world finally put a public price tag on a frontier AI lab, only to find the price barely reflected the lab itself.

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, known internationally as THUDM, had been studying knowledge graphs and language models long before they became popular. In 2019, professors Tang Jie and Li Juanzi spun off this work to form the company. The architecture they brought with them was called GLM, the General Language Model, which became both the company's technical identity and the source of its name.

This origin gave them two things, but only one is frequently written about.

First, technology. In March 2023, when most Chinese AI companies had yet to release anything developers could use, Zhipu released ChatGLM-6B, an open-source dialogue model small enough to run inference on a single consumer-grade GPU. It became one of the most downloaded models that year and the first widely available Chinese instruction-tuned LLM. Hobbyists fine-tuned it on laptops, university labs used it in courses, and companies tore it apart to study its principles. The habit of releasing models for free was there from the start, and for a reason that was far from romantic: free release was how an academic spin-out got noticed.

The second thing was trust, and trust later became the business itself. Zhipu became one of the "Six Little Dragons"—the group of Chinese LLM startups that emerged during the generative AI wave. Before listing, it assembled an unusually broad investor list: Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Hillhouse Capital, Qiming Venture Partners, Chinese local government funds, and Saudi Aramco's Prosperity7 Ventures, totaling about 1.5 billion USD. For a Tsinghua spin-out with state capital on its shareholder roster, Chinese state-owned banks could confidently procure from it; no one on the procurement chain would ever have to defend the decision. This access wasn't a soft advantage. As the revenue structure reveals, it was the entire business engine.

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

What Exactly Does Zhipu Sell?

In 2025, out of Zhipu's 724 million RMB in revenue, 534 million came from privatization deployments—73.7%.

This single number reconstructs the entire company.

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

Clients are Chinese state-owned enterprises, banks, and government agencies. These entities cannot put sensitive data on someone else's cloud, nor would they buy foreign models at any price. For them, a Tsinghua-backed, state-invested, domestically-deployed model isn't just one option among many; it's the only option. This is the Tsinghua pedigree cashed in and converted into invoices.

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

But the company's shape is dictated 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 privatization deployment segment itself dropped from 66.0% to 48.8%. The business grows, but margins shrink because growth means hiring more engineers to sit in more buildings. There is no version of privatization deployment where serving the tenth client is cheaper than serving the first. Headcount has no cost curve, just more heads.

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

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

Put these two lines together. Between 2022 and 2024, Zhipu's revenue grew roughly 5 times. Over the same three years, its computation bill grew over a hundred times. And in the first 6 months of 2025, the company spent more on computation alone than it earned in all of 2024.

These two lines are not converging.

Traditional software, written once and copied at zero cost, makes software companies the most profitable businesses in history. Large language models break this. They are written once, but every time someone uses them, the company pays again. Revenue grows linearly; computation consumption grows on a curve, and that curve steepens as context windows lengthen and reasoning chains run longer. In 2025, for every 1 RMB Zhipu earned, it paid far more than 1 RMB to chip makers and cloud providers.

The Open-Source Paradox

Zhipu's best 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 them on your own hardware, modify them, build a product, sell it—never pay Zhipu a cent. This isn't a stripped-down community edition; this is the flagship, the one the company uses to benchmark against frontier US 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. It comes from selling deployment, integration, and service. The open-source weights are marketing. The privatization contracts are the product. Releasing the model for free doesn't cannibalize revenue because the paying customers never intended to rent the model via API anyway. They were always going to pay someone to come install it in their building.

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

The proof that the strategy works isn't downloads—downloads are easy to juice and hard to monetize. The proof is price. While domestic peers were slashing prices to grab market share, Zhipu raised its API pricing by 83%, and demand still exceeded supply. The open platform's annual recurring revenue reached 1.7 billion RMB, about 240 million USD, up 60 times year-over-year—a figure Zhang Peng gave during 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 can't touch.

Changing the demand curve are agents. The GLM-5 series is optimized for long-horizon software engineering tasks. The model can work continuously for hundreds of iterations rather than stopping after one answer. Zhipu's programming packages plug into the tools developers already use. When a code agent runs autonomously for an hour, it consumes not one query's worth of tokens, but thousands of times that. Chairman Liu Debing's argument is that the resulting usage and price growth is durable, not a peak, because it's driven by models getting better and users letting them do more work.

This is the one-sentence version of the bull case: agents are token furnaces, and Zhipu sells tokens.

The Float Machine

But none of this explains a stock that went from 116.20 HKD in January to an intraday high of 2980 HKD on June 22—a 25-fold increase in less than six months. To explain that, you have to look at the plumbing.

On January 8, 2026, Zhipu listed at an IPO price of 116.20 HKD, selling approximately 43 million shares (including the overallotment), representing roughly 9.65% of its share capital. Of these, 11 cornerstone investors took about 2.98 billion HKD worth, close to 70% of the shares offered. Cornerstone investors are large institutions brought in before a Hong Kong issuer lists: they commit to buying big blocks, guaranteed an allocation, and in exchange agree not to sell for six months. Retail investors oversubscribed the remaining portion by over a thousand times.

Do the subtraction. On the first day, the shares that could actually trade were roughly 17.35 million. Less than 4% of the company.

A stock with a 4% float doesn't have the price of a normal stock. It has a clearing level between a tiny number of willing sellers and any demand that appears. And what appeared in the first half of 2026 was every investor on earth who wanted exposure to Chinese frontier AI and had no pure-play public stock to buy before January. It wasn't DeepSeek—that's private. It wasn't Moonshot AI—that's private. It wasn't Huawei—that doesn't list or sell models. It was Zhipu, and MiniMax listed a day later. That was the entire menu.

UBS put it plainly: the valuation reflected a scarcity premium and a limited number of tradable shares. Bloomberg later observed that Zhipu's stock was Asia's most volatile, largely because there were so few shares in circulation.

The machine ran for six months. Then July came and ran it in front of everyone.

On July 2, as the cornerstone lock-up period neared, the stock fell nearly 17% in a single day. Nothing happened; it was just the anticipated supply on the horizon. Closing at 1754 HKD, the market cap fell below 800 billion. The trade was too crowded and the float too thin; the mere prospect of 25.68 million shares, 5.8% of the company, becoming tradable erased roughly a sixth of the market cap in a single day.

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

Within 24 hours, Zhipu issued shares into the rally. It placed approximately 19.8 million new shares at 1588 HKD each, raising about 31.4 billion HKD, slightly over 4 billion USD. It was one of Hong Kong's largest placements of the year, six times the size of Zhipu's own IPO. CICC and CCB International were the bookrunners. The share price was priced at a discount of about 13% to the previous close—the discount needed to get institutions to take the stock at that level.

After the lock-up release and the placement, only about 13.5% of Zhipu's issued shares were freely tradable.

The company didn't fix the float. It monetized the float.

The control experiment ran the very next day. MiniMax, another Chinese model developer listed in Hong Kong in January, saw its own lock-up period expire on July 9. The founder extended the lock-up, strategic shareholders promised not to sell. It still fell more than 20% intraday.

The difference wasn't in the underlying architecture—the architecture was largely the same. The difference was that MiniMax tried to raise its M3 model's price, the market refused, and it had to cut prices. Zhipu raised prices by 83% and the market took it. A thin float amplifies any prevailing market conviction, but it doesn't create conviction out of thin air. When conviction exists, a thin float can turn a good quarter into a 25x gain. When conviction doesn't exist, the same thin float can turn an unlock into a rout.

Training Without Nvidia

For anyone tracking the Chinese AI tech stack, the most significant statement Zhipu made was made quietly.

The open-source flagship model GLM-5, released in February 2026, was reportedly trained and deployed on Chinese accelerators, not Nvidia hardware: Huawei's Ascend, as well as chips from Cambricon, Moore Threads, and Kunlun Core. On the earnings call, Zhang Peng said Zhipu had been accelerating its use of domestic chips since February to meet the sharp rise in computing demand. The R&D budget included co-design work for adapting to domestic chips. 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.

Zhipu chose this path less by active decision and more by being chosen. In January 2025, the