一个老外眼裡的智譜AI：模型免費、巨額虧損，市值為何一度超過美團？

Original Author: Robonaissance

Original Translation: TechFlow

Foreword: Zhipu AI's Hong Kong-listed shares surged 25-fold within six months, briefly surpassing Meituan's market cap. However, its 2025 revenue was only 724 million RMB, with a loss of 4.72 billion RMB. Its flagship model, GLM-5.2, is freely available for download under the MIT open-source license. This isn't a crazed market; it's pricing in scarcity, sovereignty, and a tiny, easily maneuvered free float. With its Tsinghua pedigree, state-backing, and 73.7% of revenue from private deployments for state-owned enterprises – this is what Zhipu is truly selling.

On July 2, 2026, the world's first publicly listed AI lab saw its stock price plummet nearly 17% in a single day. Six days later, on the morning of July 8th, as the lock-up period expired and approximately 46 billion Hong Kong dollars in frozen shares were unlocked, the stock price instead rose by 13%. Within 24 hours, the company took advantage of the rally to issue 4 billion USD in new shares.

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

The underlying business is more fantastical than the candlestick charts. In 2025, Zhipu's revenue was 724 million RMB (~$105 million USD). Its loss was 4.72 billion RMB (~$650 million USD). R&D expenditure was 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 single cent.

At the end of June, the market valued the company at 1 trillion Hong Kong dollars, around $128 billion USD. That's higher than Meituan – a company that delivers food to hundreds of millions of people and is genuinely profitable.

A simplistic 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 free float so small it can be easily pushed around. This is the story of these three elements, and what happens when the world finally gives a frontier AI lab a public price tag, only to find that 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, and this difference explains most of the company's subsequent shape.

The Tsinghua University Knowledge Engineering Group, known internationally as THUDM, had been researching knowledge graphs and language models long before they became popular. In 2019, professors Tang Jie and Li Juanzi spun off the 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 brought two things, though only one is frequently written about.

First, the technology. In March 2023, while most Chinese AI companies hadn't yet released anything developers could use, Zhipu released ChatGLM-6B, an open-source dialog 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 for courses, and companies took it apart to study its principles. The habit of releasing models for free existed from the start, for a reason that is anything but romantic: free release was how an academic spin-out got noticed.

The second thing was trust, which later became the business itself. Zhipu became one of the "Six Little Dragons," the cohort of Chinese LLM startups emerging from 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, Chinese local government funds, and Saudi Aramco's Prosperity7 Ventures, totaling approximately $1.5 billion. A Tsinghua spin-out with state-backed investors on its cap table meant Chinese state-owned banks could confidently purchase its services; no one on the procurement chain needed to defend the decision. This access isn't a soft advantage. Looking at the revenue structure makes it clear: this 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. The Chairman is Liu Debing. For a company once valued higher than Meituan, this is a very small building filled with academics.

What is Zhipu Actually Selling?

In 2025, out of Zhipu's 724 million RMB revenue, 534 million came from private deployment – 73.7%.

This single number redefines the entire company.

Private deployment means Zhipu's engineers go to the client's premises, 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 client 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 will not buy foreign models at any price. For them, a model with a Tsinghua pedigree, state investment, and domestic deployment isn't just one option – it's the only option. This is what the Tsinghua pedigree buys, converted into invoices.

The remaining 190 million RMB, or 26.3%, came from 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, with gross margins climbing from 3.3% to 18.9%, as inference optimization and scale drive down the marginal cost per token.

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 scales, margins shrink, because scaling means hiring more engineers to sit in more buildings. There's no version of private deployment where serving the tenth client costs less than serving the first. Labor doesn't have a cost curve; it only has headcount.

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

The problem lies below, in the other big number that determines everything. Zhipu's computing power costs paid to third parties were 14.6 million in 2022, 311.7 million in 2023, and 1.55 billion in 2024. In just the first half of 2025, according to its prospectus, 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 computing bill grew 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, written once and copied at zero cost, makes software companies the most profitable businesses in history. Large models break this. You write it once, but every time someone uses it, you pay again. Revenue grows linearly, but computing consumption grows on a curve, and that curve steepens as context windows lengthen and inference chains run longer. In 2025, for every 1 RMB Zhipu earned, it paid far more than 1 RMB to chip manufacturers and cloud providers.

The Open-Source Paradox

Zhipu's best model is free.

GLM-5.2, released in mid-June 2026, supports a maximum context window of 1 million tokens and is released under the MIT open-source license. Download the weights, run them on your own hardware, modify them, build products, sell them – 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 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-sourcing the weights is marketing; the private deployment contract is the product. Giving away the model doesn't cannibalize revenue because the paying customers never intended to rent the model via API anyway. They always intended to pay someone to come and install one on their premises.

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

The evidence the strategy works isn't download counts – those are easily faked and hard to monetize. The evidence is price. While Chinese peers were slashing prices for market share, Zhipu raised its API prices by 83%, and demand still exceeded supply. The open platform's ARR reached 1.7 billion RMB, about $240 million USD, up 60x year-over-year – a figure Chairman Zhang Peng gave during the company's first earnings call as a public company.

A company that can raise prices during a price war possesses something the price war cannot reach.

The factor changing the demand curve is agents. The GLM-5 series is optimized for long-horizon software engineering tasks; the model can work for hundreds of iterations instead of stopping after answering one question. The programming packages Zhipu sells plug into developer tools they already use. When a coding agent runs autonomously for an hour, it doesn't consume the tokens of a single query; it consumes thousands of times that. Chairman Liu Debing's argument is that the resulting volume and price increases are persistent, not a spike, driven by models getting better and users letting them do more work.

This is the bull case in one sentence: agents are a token furnace, and Zhipu sells tokens.

The Free 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 22nd – a 25-fold increase in less than six months. To explain that, you have to look at the plumbing.

Zhipu listed on January 8, 2026, at an issue price of 116.20 HKD. It sold about 43 million shares, including the over-allotment option, representing roughly 9.65% of its total share capital. Eleven cornerstone investors took up nearly 2.98 billion HKD, about 70% of the offered shares. Cornerstone investors are large institutions brought in before a Hong Kong issuer's IPO: they promise to buy a large chunk, guarantee allocation, and in exchange agree not to sell for six months. Retail investors oversubscribed the remaining portion over a thousand times.

Do the subtraction. On the first day, the shares truly available for trading were about 17.35 million. Less than 4% of the company.

A stock with a 4% free float doesn't have a normal stock price. It only has a clearing level between the very small number of people willing to sell and whatever demand appears. And the demand that appeared in the first half of 2026 came from every investor on earth who wanted exposure to Chinese frontier AI and, before January, had no publicly traded pure-play. Not DeepSeek – that's private. Not Moonshot AI – that's private. Not Huawei – not listed and doesn't sell models. Just Zhipu and MiniMax, listed a day later. That was the entire menu.

UBS put it bluntly: 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 of the tiny free float.

This machine ran for six months. Then July came, and it ran the experiment in public.

On July 2nd, as the cornerstone lockup neared, the stock fell nearly 17% in a single day. Nothing happened except the anticipation of upcoming supply. It closed at 1754 HKD, market cap dropping below 800 billion. Trading was too crowded and the float too thin; the mere prospect of 25.68 million shares – just 5.8% of the company – getting unlocked erased roughly a sixth of the market cap in one day.

Then on July 7th, the lockup period expired. The cornerstone investors didn't sell. Nearly 70% of them committed to continuing their holdings. On July 8th, the stock rose 13.35%, adding over 100 billion HKD in market cap in a single day, because the decision not to sell 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 1588 HKD per share, raising about 31.4 billion HKD, slightly over $4 billion USD. This was one of the largest placings in Hong Kong that year, more than six times the size of Zhipu's own IPO. CICC and CCB International acted as bookrunners. The placement price was roughly a 13% discount to the previous close – the discount needed to get institutions to take the shares at that level.

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

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

The controlled experiment ran the next day. MiniMax, another Chinese model developer who listed in Hong Kong in January, faced its own lockup expiry on July 9th. The founder extended his lockup; strategic shareholders committed not to sell. The stock still fell over 20% during the session.

The difference isn't in the underlying architecture – the architecture is broadly similar. The difference is that MiniMax tried to raise the price of its M3 model, the market rejected it, and it had to cut prices. Zhipu raised prices by 83%, and the market accepted it. A thin float amplifies whatever market belief already exists, but it doesn't create it from scratch. When the belief is there, a thin float can turn a good quarter into a 25x rally. When the belief is absent, that same thin float can turn an unlock into a rout.

Training Without Nvidia

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

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

Zhipu didn't so much choose this path as it was chosen by it. In January