The leading players in large models are sucking up the primary market.
A financing frenzy, defined by industry insiders as the "eve of a market purge," is taking place in the global large - model race.
Before May even ended, three rounds of financing totaling over $7 billion flooded the Chinese market: Kimi completed a financing of approximately $2 billion at the beginning of the month; Step - by - Step Star was reported to be close to completing a financing of nearly $2.5 billion; and after DeepSeek received external financing for the first time, its post - investment valuation range was pushed to between $45 billion and $50 billion.
In the European and American markets, OpenAI, Anthropic, and SpaceX (after its merger with xAI under Elon Musk) are expected to go public this year. The combined valuation of the three companies exceeds $3 trillion.
This torrent of capital across the Pacific is surging towards the last leading players in the large - model race at an unprecedented speed and scale. It should be noted that not all companies can secure funding. On the contrary, for the vast majority of companies, the music has stopped.
However, for the few companies that can obtain funding, this may be their last chance to reach the next level.
01
The "Musical Chairs" Game Enters the Final Stage
Since the beginning of this year, the Chinese large - model race has first presented two significant reports to the capital market.
First, on January 8th, Zhipu AI, which had been established for six years, officially listed on the Hong Kong Stock Exchange. With an issue price of HK$1.16 billion per share, it won the title of the "world's first large - model stock." On the first day of listing, its stock price closed up 13.17%, with a market value of HK$57.9 billion.
Just one day later, MiniMax, founded in early 2022, also listed on the Hong Kong Stock Exchange. Its stock price soared 109.09% on the first day, and its market value exceeded HK$100 billion, setting a new record for the fastest time from establishment to listing for a global AI company.
Moreover, after the listing of these two companies, their stock prices continued to rise. As of May 15th, Zhipu AI's stock price rose from the issue price of HK$116.2 to a maximum of HK$1,229, more than ten times the increase in four months after listing. MiniMax also showed a nearly vertical growth curve.
JPMorgan Chase maintained the "Overweight" rating for these two companies in a recent research report but also gave a sober judgment: The market's valuation of them already implies the prospect that Zhipu AI's Annual Recurring Revenue (ARR) will reach $1 billion and MiniMax's will reach $700 million by the end of 2026.
The frenzy in the secondary market quickly spread to the primary market.
On May 6th, Kimi was reported to be about to complete a new round of financing of approximately $2 billion, with a post - investment valuation exceeding $20 billion. This round of financing was led by Meituan Longzhu, with participation from China Mobile, CPE (CITIC Industrial Investment Fund), etc. Longzhu invested more than $200 million.
Counting the three rounds of financing since the end of last year, Kimi has raised more than $3.9 billion in half a year, with a total financing of over 37.6 billion yuan, becoming the domestic large - model startup with the highest cumulative financing.
Another star company, DeepSeek, which caused a stir in the global AI circle in 2025 with its DeepSeek - R1 model, had always adhered to the principle of "not introducing external funds." However, the situation changed this spring.
According to a report by The Wall Street Journal on May 7th, DeepSeek is raising billions of dollars from government - supported investors, and the National Artificial Intelligence Industry Investment Fund is in in - depth negotiations to participate in the financing.
People familiar with the matter revealed that Liang Wenfeng himself even plans to invest 20 billion yuan personally. According to industry estimates, the post - investment valuation is expected to exceed $50 billion.
In addition, Step - by - Step Star was reported to be about to complete a financing of nearly $2.5 billion. It has dismantled its red - chip structure and is fully sprinting for a Hong Kong IPO. The list of its investors includes consumer electronics industry chain enterprises such as Huaqin Technology, Longcheer Technology, and ZTE.
Shuanshu Technology completed two large - scale financings in 2026: over 600 million yuan in Series A+ and nearly 2 billion yuan in Series B, with a cumulative financing of nearly 2.6 billion yuan in less than four months.
AI native infrastructure service provider Wuwen Xinqiong also officially announced the completion of a Series B financing of over 700 million yuan on May 7th.
If we shift our focus from China to the other side of the Pacific, the protagonists of this capital feast are even more massive.
According to current public information, SpaceX is scheduled to list on the NASDAQ in June, with a target valuation of $1.75 trillion. If it succeeds, it will surpass Saudi Aramco to become the largest IPO in human history. OpenAI plans to go public in the fourth quarter, with a valuation of approximately $852 billion. Anthropic also plans to complete its listing this year, and the secondary market values it at over $1 trillion.
In just the primary - market financings completed in February and March, OpenAI and Anthropic each received hundreds of billions of dollars in funding. The combined valuation of the three giants exceeds $3 trillion, far exceeding any previous technology IPO combination.
The core fact outlined by these running numbers is that capital is concentrating on a very small number of leading players in the race at an irreversible speed.
Looking back, during the "Hundred - Model Battle" in 2023, hundreds of startups competed on the same stage. However, in 2025, some media reported that AI model - layer companies only completed 22 rounds of financing throughout the year, with a total disclosed amount of 9.4 billion yuan. The proportion of large - model financing in total AI investment plummeted from 51% in 2024 to 14%, indicating that the industry's "battle royale" has eliminated more than 90% of the players.
However, when more than $7 billion flowed into three leading companies within three days in May 2026, the signal from the industry was very clear: Capital is no longer "transfusing" the entire industry but "filling up" the last few players.
02
The Economics of the Token Factory
This capital frenzy did not arise out of thin air. It is driven by both the transformation of technological routes and the reshaping of market logic. To understand this frenzy, we need to examine it from both internal and external perspectives.
The industry narrative has undergone a fundamental shift in the past year.
Before 2024, the core story of large models was "who has more parameters and who is smarter." Major manufacturers competed to burn money to train models, vying for the upper limit of intelligence.
However, the emergence of long - range agents such as OpenClaw (commonly known as "Lobster") in February 2026 completely opened the "Pandora's box" of computing power consumption. An agent needs to call the model dozens or even hundreds of times to handle a complex task, and the amount of Tokens consumed has soared from a few thousand in traditional single - round conversations to hundreds of thousands or even millions.
Since then, the industry has stopped competing for the "upper limit of intelligence" of models. Instead, it is competing to produce a large number of Tokens at a lower cost and in a more stable way. As Huang Renxun, the founder of NVIDIA, defined in the "Economics of the Token Factory," this is an industrial revolution driven by the explosion of real - world demand, the imbalance of supply - demand structure, and global computing power competition.
The data from the National Data Bureau clearly shows how "brutal" this explosion is. The average daily Token calls in China soared from 100 billion at the beginning of 2024 to 140 trillion in March 2026, an increase of more than 1,000 times in two years.
Since 2026, the cumulative increase of the A - share AI computing power sector has exceeded 55%. The monthly revenue of leading large - model enterprises has exceeded 1 billion yuan, and the revenue of some enterprises in 20 days has already exceeded the full - year scale of 2025.
The structural imbalance on the supply side has caused the pricing power of Tokens to shift sharply upstream.
The High - Bandwidth Memory (HBM) market is monopolized by Samsung, SK Hynix, and Micron. The production expansion cycle is as long as 24 to 36 months, resulting in a HBM shortage of more than 40% in 2026. Electricity costs account for more than 60% of the Token production cost, and the power infrastructure construction cycle for large - scale data centers is as long as 3 to 5 years.
This actually leads to a "first - principle logic" that determines the direction of the large - model industry today - the current large models are no longer just software but a hybrid of "software + cloud computing + heavy - asset industries." Every time a user chats, searches, or gets a response, GPUs and electricity are being consumed in real - time.
When the "marginal cost" of the model no longer approaches zero, whoever controls the most computing power resources and can produce Tokens at the lowest cost will have the pricing power. The competition for these resources is not about algorithms but real money.
On a macro level, the huge investments of international technology giants in AI infrastructure have also intensified the industry's focus on the current competition.
According to the latest capital expenditure guidelines announced by various companies during the earnings season in April 2026, Microsoft's annual AI infrastructure capital expenditure is expected to reach $190 billion; Alphabet has raised its annual capital expenditure forecast to between $180 billion and $190 billion, further increasing the guidance from February; Meta also raised its forecast to between $125 billion and $145 billion when releasing its earnings report on April 29th, citing rising component prices and increasing data - center construction costs as the reasons for the increase; Amazon maintained its forecast at approximately $200 billion.
Calculated based on the upper limit of the guidelines, the total capital expenditure of the four giants in 2026 is approximately $725 billion. Obviously, this is not just an expenditure for an industry but the completion of a power - supply system for a new intelligent era and the laying of a computing - power "power grid" for all AI applications.
On the other hand, the listing - window effect brought about by some startups has also accelerated the financing rhythm of the primary markets of Chinese and American VCs. In particular, the sharp rise of Zhipu AI and MiniMax after their listings has established a reference benchmark for "how much a large - model company is worth" in the secondary market. This has stimulated the anxiety of other unlisted companies about the future. If they do not complete their valuation during this window, once the market gets tired and the valuation corrects.
As a result, Step - by - Step Star completed all the steps from dismantling the red - chip structure, restructuring into a joint - stock company, to sprinting for a Hong Kong IPO within a few months; and Kimi's valuation soared from approximately $4.3 billion to over $20 billion, reflecting both the improvement of its fundamentals and the acceleration of capital competing for a stake in the "next listed company."
03
Determining the Future Victory
On one hand, there is a capital frenzy, and on the other hand, the focus of competition is shifting. The industry generally believes that future competition will mainly focus on three aspects.
First, commercialization and monetization will become the "top priority" for each company.
It must be recognized that a fundamental change is taking place in the large - model industry in 2026, that is, the "AGI premium" is cooling down.
In the past two years, there has been a key implicit premise for the high valuations of AI companies in the capital market: the Scaling Law continues to be effective, and the model's capabilities will rapidly improve with the input of computing power. AGI is just a matter of time. Investors are willing to accept short - term losses and discount the "future efficiency revolution" into today's stock prices.
However, in 2026, although AI is still progressing, the form of progress does not seem as consistent as before. OpenAI revised its principle document and reduced its direct references to AGI; Demis Hassabis of DeepMind also publicly admitted that the current system still has obvious gaps in continuous learning and long - term planning.
At this time, the market's focus has shifted from "who is closer to AGI?" to "who can get customers to pay? Who can reduce the inference cost? Who can survive?"
In fact, the commercialization signals sent by some leading manufacturers are very clear. ByteDance's Doubao, which has long had 345 million monthly active users with a free model, quietly launched a paid plan of up to 5,088 yuan per year on the Apple App Store not long ago. OpenAI has significantly strengthened the paid enterprise capabilities of Codex and actively restricted the top - level use of free users.
This marks that the global large - model industry has entered a rational and mature stage from burning money to gain traffic. The core proposition of the competition has shifted from "whose model is stronger" to "whose model can make money first."
Second, computing power cost has become the ultimate KPI.
With the development of the large - model industry, in the foreseeable future, inference ability, long - text processing, and multi - modality will no longer be moats. After DeepSeek V4 brought the open - source model close to the level of GPT - 4, the industry realized for the first time systematically that the model's capabilities themselves are easier to catch up with than expected.
As models are gradually becoming "commoditized," the capital market is starting to ask: What else do you have besides the model?
This has led to a shift in the industry narrative.
In 2023, companies competed on "more parameters and longer context." Today, companies are talking about which terminals they have locked in, which industrial chains they have bound, and which user entrances they have mastered.
JPMorgan Chase pointed out in a research report that the market's valuation of Zhipu AI already implies an expected ARR of approximately $1 billion by the end of 2026. Under the new evaluation framework, judging the value of a company is no longer just based on evaluation scores but on who its customers are, whether its cash flow is healthy, how many paid scenarios it has opened up, and how irreplaceable it is among its partners.
Third, the explosion of agents and the differentiation of paths.
2026 is generally regarded as the first year of the explosion of agents in the industry. When we focus on the quantity and speed of agents released by manufacturers, what is more worthy of attention is the future differentiation between the ToB and ToC paths.
One path is along the direction of "embedding into the production process," betting on the improvement of certain production efficiency. The other path leads to real - life scenarios for individuals, betting on user mindsets and long - term scale.
There is no right or wrong between the two paths, but their capital consumption rhythms and requirements for business - model maturity are completely different. Serving enterprise customers requires building an iron triangle among reliability, integration, and security, which is a long - term trust - building process. The C - end scenario relies on the self - reinforcement of the data flywheel and user mindsets, burning money in the early stage but having a strong scale effect in the later stage.
In the context of high computing - power bills and a higher concentration of financing, whether a company can achieve a closed - loop and positive cash flow in its own track will directly determine the ranking after the "eve of the market purge" in 2026.
04
Conclusion
For today's investors, it is no longer a multiple - choice question of "which direction to invest in" but a reshuffle game of making "end - game bets" on a limited number of leading players. The three variables of technological route, scenario selection, and capital endurance will jointly determine who can stay at the table and who will be asked to leave.
In an era when models are becoming increasingly commoditized, the real decisive factor may no longer be just the technological ability itself but how to turn technological ability into services that customers are willing to pay for continuously, turn computing - power input into verifiable output, and turn a product into a healthy company.