Zhipu AI's valuation exceeds one trillion yuan, China's large models put an end to the "cheap alternative" narrative

On June 18th, Elon Musk made an assertion on social media: Chinese large models won't truly catch up with the world's top - level models until the first quarter of 2027. When Tang Jie, the co - founder of Zhipu, saw this, he simply replied with four words: "Won't take that long."

Musk then added: "Maybe it's true in benchmark tests. But if judged by the practical value of real - world implementation, even if they catch up in the first quarter of next year, it would be very impressive." Marc Andreessen, the co - founder of a16z, reposted that conversation and only added one word: "Interesting."

There is much debate outside about whose judgment is more accurate. But when looking at the events before and after in conjunction with this conversation, it's quite interesting: Three days earlier, on June 15th, Zhipu's stock price soared by 32.82% in a single day. On the 17th, GLM - 5.2 was officially open - sourced and launched. On the 19th and 22nd, Zhipu's intraday stock price soared to HK$1.3 trillion, entering the trillion - club.

Has the "practical implementation value" been achieved? This question is still under debate. But Zhipu's market value has increased from HK$52.8 billion at the time of its listing in January to trillions in less than half a year, a rise of 2000% - the market has already spoken with this figure.

01

From a Substitute to a Competitor

On January 8th, 2026, Zhipu was listed on the Hong Kong Stock Exchange at an issue price of HK$116.2. At that time, the market positioned it as the "Chinese version of OpenAI" - this title has positive connotations, but there is also an underlying meaning: a "substitute" and a "chaser" of OpenAI.

The substitute narrative was almost the collective situation of the entire Chinese large - model industry at that time.

The change occurred five months later. On June 17th, Zhipu released GLM - 5.2. It is under the MIT license and completely open - sourced. It achieved a score of 74.4 on the programming benchmark FrontierSWE. For comparison, Anthropic Opus 4.8 scored 75.1, and GPT - 5.5 scored 72.6.

This is the first time a domestic open - source model has entered the same competitive arena as the world's top - level closed - source models. The AI research institution Proximal wrote in its evaluation report that GLM - 5.2 is "the first model that truly narrows the huge technological gap between Anthropic/OpenAI and other model providers." American open - source researcher Nathan Lambert was more direct in his evaluation: GLM - 5.2 surpasses Google Gemini in terms of agent capabilities. On the comprehensive list of Artificial Analysis, GLM - 5.2 ranked between GPT - 5.5 and Opus 4.8 with 51 points, being the highest - ranked open - source model to date.

On the same day, Zhipu's guidance status on the Science and Technology Innovation Board was changed to "guidance acceptance." It only took 11 days from the completion of guidance filing on June 6th to this stage. Both technological breakthroughs and the reconstruction of the capital narrative were achieved simultaneously.

However, the catalyst was actually laid earlier. On June 12th, the U.S. Department of Commerce cited the Export Administration Regulations and required Anthropic to obtain a license to provide access to its cutting - edge models to foreign users. Anthropic immediately closed the global access to its two flagship models. Suddenly, a wall was erected, and open - source models caught the overflowing demand. After the sharp rise on June 15th, Zhipu's stock price rose another 26.14% on the 18th. On June 22nd, the intraday stock price soared to HK$2980, corresponding to a market value of HK$1.3 trillion, equivalent to three Meituan or four JD.com.

The term "substitute" no longer exists. A model that can compete with Anthropic and OpenAI on the same benchmark list can hardly be called a "follower" of anyone. In less than half a year, the competitive narrative has replaced the substitute narrative.

02

The Logic of Valuation

Of course, there is an unavoidable numerical question here: Zhipu's annual revenue in 2025 was 724 million yuan, and it had a loss of 4.718 billion yuan. What exactly supports this trillion - level market value?

The answer may be found in a growth curve: In Q1 of 2026, Zhipu's API ARR (Annual Recurring Revenue) reached 1.7 billion yuan. A year ago, this figure was 28 million yuan. Three months ago, it was 500 million yuan. More importantly, this 1.7 billion was achieved after an 83% price increase. The price went up, but the call volume increased by 400%. After nearly two years of price wars in the large - model industry, Zhipu completed a reverse operation - it is proving that the competitiveness of a model will ultimately be reflected in pricing power. JPMorgan's judgment is in line with this: In the AI era, the ability to continuously raise prices is a better indicator of model competitiveness than benchmark scores.

The change in ARR has also brought about a switch in the valuation framework. Zhang Youyu, a specially - appointed researcher at Peking University, summarized the logic of this round of rise as a fundamental switch in the valuation anchor - from the "research institution" framework (with the core narrative being R & D investment) at the initial stage of listing to the "SaaS valuation" framework after having measurable cash - flow indicators. This logic is not unfamiliar in Silicon Valley. Anthropic's ARR was about $9 billion at the end of 2025. By the time it completed its Series H financing in May 2026, this figure had risen to $47 billion - it quadrupled in five months, and the post - investment valuation rose to $965 billion, surpassing OpenAI for the first time and becoming the world's most highly - valued AI company. The premium given by the market is a bet that the growth curve is not over yet. Zhipu's logic is in the same vein, but the curve is steeper.

There is also a detail that can confirm this logic from the side. The API pricing for GLM - 5.2's long - context output is $4.13 per million tokens, while that of the similar Fable 5 output is about $50. This price difference is a safety cushion for Zhipu's price increase - because the base price is low enough, it has room to raise the price without fearing user loss. The price increase itself has become a bargaining chip rather than a risk.

Against this background, it is not difficult to understand the use of the 15 billion yuan raised on the Science and Technology Innovation Board. 12 billion is invested in the R & D of the large - model base, 2 billion is used for the MaaS service platform, and 1 billion is used to supplement working capital. It is listed on the Hong Kong Stock Exchange, and the listing on the Science and Technology Innovation Board is also on the way. By advancing on two fronts, it aims to anchor the narrative position of the "representative of Chinese AI" in both markets - and to reserve enough resources in advance for the arms race of the next - generation models.

03

Not Just Zhipu

Zhipu's trillion - level market value is a capital event, but it reflects the changes in the entire Chinese large - model industry.

In May this year, Dario, the CEO of Anthropic, released a report. The core conclusion is simple: The United States must lock in its leading edge over China in large models for 12 to 24 months - as long as it can maintain this line, the United States can retain its dominance.

Then, this judgment lasted for less than a month.

According to the comprehensive data of Artificial Analysis, in 2023, the comprehensive ability of China's cutting - edge large models was about 60% of that of the top U.S. models; by 2025, this figure had risen significantly to 90%. The AI Index Report of Stanford University believes that the gap has been narrowed to 0.3% - 1.7%. Among the top 20 in the global model comprehensive ability list, except for a French company, Mistral, all the others are from the United States or China. Musk believes that the time gap between China and the United States is 10 months, while the general assessment in the current international technology community is that the time gap between Chinese and U.S. large models has been shortened to about 7 months.

There is an underlying structure to this pursuit that is not always noticed by the outside world: the collective acceleration of the open - source ecosystem.

There is a unique phenomenon in the Chinese large - model industry - companies are "using each other's strength." Kimi K2 adopted the MLA potential attention design verified by DeepSeek in its architecture; DeepSeek V4 introduced Kimi's Muon optimizer, reducing the proportion of activated parameters from 5% to 3%, while the AI intelligence index jumped from 42 to 52; both GLM and Qwen adopted the GRPO reinforcement learning scheme verified by DeepSeek. The technical paths released by a company in the open - source community often appear in the next versions of other companies within a few months.

This is almost impossible in the U.S. closed - source camp. OpenAI and Anthropic each guard their own technological barriers, do not share with each other, and have no incentive to share. In contrast, in China, the latest verified and effective paths quickly spread throughout the industry and become public infrastructure - every technological breakthrough is not only the gain of a single company but also a leap for the entire ecosystem.

More notably, this kind of collaboration occurs against the background of a huge disparity in R & D investment. According to public data, the combined R & D investment of Zhipu and MiniMax is about one - tenth of that of Anthropic. Achieving similar results with fewer resources, open - source collaboration is a lever that cannot be ignored.

The results brought about by this effect can already be quantified. DeepSeek's platform usage ranked first globally in May; MiniMax's overall gross profit margin in 2025 was 25.4%; Goldman Sachs' research shows that the combination of Kimi K2.6 and DeepSeek V4 Pro has narrowed the overall performance gap to less than 1% of the top - level Fable 5 at about half the cost.

The gap is narrowing, and the speed of narrowing is also accelerating.

Conclusion

This is not just the story of one company, although Zhipu's trillion - level market value makes it seem like the protagonist. But behind the scenes, a structural change is taking place - the Chinese large - model industry is changing from a follower to a competitor that can be compared on the same list. This change can be traced in technology, has traces in business, and has been priced in the capital market.

The technological window is narrowing, the spread of open - source is accelerating, and the inflection point of commercialization has appeared earlier than any prediction. These events together do not seem to be the result of the efforts of a single company but rather a change in the gravitational force of the entire industry.

So Tang Jie's words "Won't take that long" are, in a sense, not just a refutation but more like a live broadcast on the field.

This article is from the WeChat official account "Manfon Slow - Motion", author: Gulu, editor: Guzi, reprinted with permission from 36Kr.