The global large model track is experiencing a financing frenzy that industry insiders have defined as the "eve of a clearance."
Even before May ended, three funding rounds totaling over $70 billion flooded into the Chinese market: Kimi raised about $20 billion at the beginning of the month, StepFun was revealed to be close to finalizing nearly $25 billion in financing, and DeepSeek's valuation range was pushed to $45-50 billion after accepting external investment for the first time.
In the European and American markets, OpenAI, Anthropic, and SpaceX (merged with xAI under Musk's control) are expected to go public within the year, with a combined valuation exceeding $3 trillion.
This torrent of capital spanning the Pacific is pouring into the final leading players in the large model track with unprecedented speed and scale. It is important to note that not all companies can get the money; on the contrary, for the vast majority, the music has stopped.
But for those few who can secure funding, this might be the last train to the next level.
01
The "Musical Chairs" Game Reaches Its Endgame
Since the beginning of the year, China's large model track has taken the lead in submitting two heavyweight answers to the capital market.
First, on January 8th, Zhipu AI, founded six years ago, officially listed on the Hong Kong Stock Exchange, seizing the title of "the world's first listed large model company" with an issue price of HKD 11.6 per share. Its stock price closed up 13.17% on the first day, reaching a market capitalization of HKD 57.9 billion.
Just one day later, MiniMax, founded in early 2022, also listed on the Hong Kong Stock Exchange. Its stock price skyrocketed 109.09% on the debut, breaking through the HKD 100 billion market cap mark and setting a new global record for the fastest journey from founding to IPO for an AI company.
Moreover, after going public, the stock prices of both companies continued to rise. As of May 15th, Zhipu's price soared from the HKD 116.2 issue price to a high of HKD 1,229, increasing more than tenfold in four months. MiniMax also charted a nearly vertical growth curve.
J.P. Morgan recently maintained an "Overweight" rating on both companies in a research report but offered a sober judgment: their valuations already imply expectations for Zhipu's annualized recurring revenue (ARR) to reach $1 billion and MiniMax's to reach $700 million by the end of 2026.
The frenzy in the secondary market quickly transmitted to the primary market.
On May 6th, it was revealed that Moonshot AI (Kimi) was about to complete a new funding round of approximately $20 billion, pushing its post-money valuation beyond $20 billion. This round was led by Meituan Longzhu, with participation from China Mobile, CPE (CITIC Private Equity), etc. Longzhu invested over $200 million.
Counting three funding rounds since late last year, Moonshot AI has raised over $3.9 billion within six months, with total funding exceeding RMB 37.6 billion, making it the AI startup with the highest cumulative financing in China.
Another star company, DeepSeek, which stirred a storm in the global AI circle in 2025 with its DeepSeek-R1 model, previously adhered to a "no external funding" path. But this spring, the wind changed.
According to a May 7th report by The Wall Street Journal, DeepSeek is raising tens of billions of dollars from government-backed investors. The National Artificial Intelligence Industry Investment Fund is in deep negotiations to participate.
Informed sources revealed that Liang Wenfeng himself even plans to invest RMB 20 billion of his own money in the subscription. Industry calculations estimate the post-investment valuation could exceed $50 billion.
Additionally, StepFun was reported to be completing nearly $2.5 billion in financing, having dismantled its VIE structure and is sprinting towards a Hong Kong IPO. Its investor lineup includes consumer electronics supply chain companies like Huaqin Technology, Longcheer Technology, and ZTE.
Shengshu Technology completed two large funding rounds in 2026: an over RMB 600 million Series A+ and a nearly RMB 2 billion Series B, accumulating nearly RMB 2.6 billion in less than four months.
AI-native infrastructure service provider Unisound also announced the completion of an over RMB 700 million Series B funding round on May 7th.
If we shift our gaze from China across the Pacific, the protagonists of this capital feast are even larger.
Based on current public information, SpaceX is set to list on Nasdaq in June with a target valuation of $1.75 trillion. If successful, it would 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 IPO within the year, with its secondary market valuation already exceeding $1 trillion.
In primary market financing rounds completed in February and March alone, OpenAI and Anthropic have each taken in hundreds of billions of dollars worth of ammunition. The combined valuation of these three giants exceeds $3 trillion, far surpassing any previous tech IPO combination.
The fundamental fact outlined by this series of racing numbers is that capital is concentrating at an irreversible speed towards an extremely small number of leading players in the track.
Looking back now, during the 2023 "Hundred Models War," hundreds of startups competed on the same stage. By 2025, media statistics indicated that AI model-layer companies completed only 22 financing rounds for the entire year, with a total disclosed amount of RMB 9.4 billion. The proportion of large model financing in total AI investment plummeted from 51% in 2024 to 14%. The industry's battle royale has already achieved over a 90% elimination rate.
However, when over $7 billion flowed into three leading companies within three days in May 2026, the signal from the industry became very clear: capital is no longer providing "blood transfusions" to the entire industry but is "filling up the tanks" of the final few contestants.
02
Token Factory Economics
This capital boom did not arise from nowhere; its background is driven by both the transformation of technology roadmaps and the reshaping of market logic. Understanding this frenzy requires perspective from both internal and external factors.
The industry narrative has undergone a fundamental shift over the past year.
Before 2024, the core story of large models was "whose parameters are larger, who is smarter." Major manufacturers raced to burn money training models, competing on the upper limit of intelligence.
But the explosion of long-context intelligent agents like OpenClaw (commonly known as "Lobster") in February 2026 completely opened the "Pandora's Box" of computing power consumption. An Agent handling a complex task requires dozens or even hundreds of model calls, with token consumption leaping from a few thousand in traditional single-turn conversations to hundreds of thousands or even millions.
Since then, the industry no longer competes on a model's "IQ ceiling" but on who can produce massive tokens at lower costs and more stably. As defined by Nvidia founder Jensen Huang's "Token Factory Economics," this is an industrial revolution driven by the explosion of real demand, supply-demand structure imbalances, and global computing power competition.
Data from the National Data Bureau clearly marks just how "brutal" this explosion is: China's daily token call volume surged from 100 billion at the beginning of 2024 to 14 trillion in March 2026, growing over 1,000-fold in two years.
Since 2026, the A-share AI computing power sector has accumulated gains of over 55%. Leading large model enterprises have seen monthly revenues exceed RMB 1 billion, with some companies' revenue in 20 days surpassing their entire 2025 annual scale.
The structural imbalance on the supply side has caused a sharp upward shift in the pricing power of tokens.
HBM (High Bandwidth Memory) is monopolized by Samsung, SK Hynix, and Micron, with expansion cycles as long as 24 to 36 months, leading to an HBM shortage of over 40% in 2026. Electricity costs constitute over 60% of token production costs, while large data center power facility construction cycles last 3 to 5 years.
This actually leads to a "first-principles logic" determining the direction of the large model industry today: large models are no longer just software but a hybrid of "software + cloud computing + heavy-asset industry." Every user chat, search, and response is backed by real-time burning of GPUs and electricity.
When a model's "marginal cost" no longer approaches zero, whoever controls the most computing resources and can produce tokens at the lowest cost holds pricing power. And the competition for these resources is not about algorithms but real money.
On a macro level, massive investments in AI infrastructure by international tech giants have also intensified the industry's focus on the current competitive points.
According to the latest capital expenditure guidance announced by companies during the April 2026 earnings season, Microsoft's full-year AI infrastructure capital expenditure is expected to reach $190 billion. Alphabet raised its full-year CapEx expectation to $180-190 billion, a further increase from February guidance. Meta also raised its expectation to $125-145 billion on April 29th during its earnings release, citing rising component prices and data center construction costs. Amazon maintained its guidance at approximately $200 billion.
Calculated at the upper limit of guidance, the combined capital expenditure of the four giants in 2026 is about $725 billion. Clearly, this is not just expenditure for one industry but the completion of the power supply system for a new intelligent era, laying the computing power "grid" for all AI applications.
On the other hand, the IPO window opened by some startups has also accelerated financing pace in the Chinese and American VC primary markets. Particularly, the skyrocketing stock prices of Zhipu AI and MiniMax after listing established a reference benchmark in the secondary market for "how much a large model company is worth." This stimulated anxiety among other unlisted companies about the future—if they don't finalize their valuation while the window is open, a market shift in taste could lead to valuation corrections.
Thus, StepFun completed all actions from dismantling its VIE structure, corporate restructuring to sprinting for a Hong Kong IPO within months. Kimi's valuation rapidly soared from about $4.3 billion to over $20 billion, reflecting both improvements in fundamentals and the acceleration of capital vying for "the next listed company."
03
Future Decisive Factors
On one side is capital fervor, on the other is a shift in competitive focus. The industry generally believes future competition will mainly concentrate in three areas.
First, commercialization will become the "top priority" for each company.
It must be recognized that a fundamental change is occurring in the large model industry in 2026: the "AGI premium" is cooling down.
Over the past two years, the high valuations granted by the capital market to AI companies contained a key implicit premise: Scaling Law remained effective, model capabilities rapidly leapfrogged with computing power investment, and AGI was just a matter of time. Investors were willing to accept short-term losses, willing to discount the "future efficiency revolution" into today's stock price.
But by 2026, AI is still advancing, but the form of progress seems less orderly than before—OpenAI revised its charter, reducing direct mentions of AGI; DeepMind's Demis Hassabis also publicly acknowledged that current systems still have significant gaps in continuous learning and long-term planning.
Now, the market's focus has shifted from "Who is closer to AGI?" to "Who can make customers pay? Who can reduce inference costs? Who can survive?"
In fact, commercialization signals from some leading players are already very clear. ByteDance's Doubao, which long held 345 million MAUs with a free model, recently quietly launched a paid plan on Apple's AppStore costing up to RMB 5,088 per year. OpenAI significantly enhanced the paid enterprise capabilities of Codex while actively limiting top-tier usage for free users.
This marks the global large model industry's transition from burning money for traffic to a rational maturity phase. The core proposition of the race has shifted from "Whose model is stronger?" to "Whose model becomes profitable first?"
Second, computing power cost becomes the ultimate KPI.
As the large model industry develops, foreseeable capabilities like inference power, long context, and multimodality will no longer be moats. After DeepSeek V4 brought open-source models close to GPT-4 levels, the industry systematically realized for the first time that model capability itself is easier to catch up with than imagined.
As models are gradually becoming "commoditized," the capital market is beginning to ask: Besides the model, what else do you have?
This has catalyzed a shift in the industry narrative.
In 2023, everyone competed on "larger parameters, longer context." Today, companies start talking about which end devices are locked in, which supply chains are bound, which user entry points are controlled.
J.P. Morgan noted in its research report that the market's valuation of Zhipu already implies expectations for its ARR to reach about $1 billion by the end of 2026. Under the new evaluation framework, judging a company's value no longer depends solely on benchmark scores, but on who its customers are, whether its cash flow is healthy, how many paid scenarios are opened, and how much irreplaceability it has built among partners.
Third, the explosion of agents and path divergence.
2026 is widely regarded in the industry as the first year of the agent explosion. While we pay attention to the number and speed of agents released by vendors, what's more noteworthy is the future divergence between the ToB and ToC paths.
One path follows the direction of "embedding into production workflows," betting on deterministic productivity improvements; the other heads towards real-life personal scenarios, betting on user mindshare and long-term scale.
There is no right or wrong between the two paths, but their demands on capital consumption pace and business model maturity are completely different. Serving enterprise customers requires forming an iron triangle between reliability, integration, and security—a long-term trust-building endeavor. C-end scenarios rely on data flywheels and the self-reinforcement of user mindshare, burning money upfront but exhibiting strong scale effects later.
Under the backdrop of high computing power bills and financing concentration reaching new heights, the ability to run closed loops and generate positive cash flow on their respective tracks will directly determine the ranking after the "eve of the clearance" in 2026.
04
Conclusion
For today's investors, it is no longer a multiple-choice question of "which direction to invest in," but a reshuffling game of making "endgame bets" on a limited number of leading players. Technology roadmap, scenario choice, and capital endurance—these three variables will jointly determine who stays at the table and who gets asked to leave.
In an era where models are increasingly commoditized, the true decisive factors may have long ceased to be just technological capability itself, but rather how to turn technological capability into services customers are willing to pay for continuously, how to turn computing power investment into verifiable output, and how to turn a product into a healthy company.
This article is from the WeChat public account "Insights & New Research" (ID: DJXYS-0309), author: Chen Wen.
