May 2026, a number completely flooded screens: the combined valuation of China's 'Four AI Dragons' exceeded 1 trillion RMB. (Source: 36Kr, May 8, 2026)
Most reports interpreted it like this: Zhipu's stock soared 7x after its HK listing, MiniMax rose nearly 4x, Moonshot AI was about to complete a new round of $2 billion financing, and DeepSeek secured investment led by the National Integrated Circuit Industry Investment Fund, anchoring its valuation at $45 billion. This is great, AI has truly exploded.
However, I believe the most noteworthy aspect this time is not the trillion-yuan figure, but the fact that behind this number, capital has already upended its own pricing logic for large model companies three times. Each time, it was convinced 'this is the one,' and each time, it was proven wrong by subsequent events.
Some might say: Capital makes mistakes, what's the big deal?
Actually, no.
Behind these three upendings lies a common underlying structure, which can be refined into a tool to help you determine where you stand in the next wave of hot money influx.
First: Pricing by Technology Narrative
Around 2022, when China's first batch of large model companies secured funding, capital's pricing logic was based on technical capability. How many parameters, benchmark scores, ability to compete with GPT-4.
A typical conversation back then went like this: 'Where does your model rank in performance evaluations?' 'Top three in Chinese leaderboards.' 'Good, here's a valuation for you.'
This logic was valid until around 2024. Because AI large models were still in the capability verification stage, technical capability itself was the scarcest asset; achieving good results meant the potential to build a moat.
DeepSeek R1 shattered this logic in early 2025.
R1 proved that similar-grade capability does not require similar-grade training costs. Creating a model with comparable effects at one-tenth the cost of US rivals directly shook the pricing premise that 'top-tier parameter count = top-tier moat.' Technical capability began transforming into a homogeneous asset, not a scarce one. (Source: Cyzone, 2026 report on the Six AI Dragons)
First time: Capital priced by 'technical score,' DeepSeek R1 pierced the myth of the technical barrier.
Second: Pricing by IPO Window
After the DeepSeek shock, capital's pricing logic made its first leap: from 'technical capability' to 'IPO certainty.'
The Hong Kong Stock Exchange relaxed market cap thresholds for listing specialized tech companies, lowering the bar for commercialized companies to 4 billion HKD. HK IPO fundraising in 2025 reached about $36 billion, a four-year high. The window was clear. Zhipu and MiniMax took the lead in filing and listing, and were wildly hyped in the secondary market. (Source: TMTpost, May 2026)
The primary market's logic switched accordingly: not betting on whose model was strongest, but betting on who could list and cash out the fastest.
The flaw in this logic was quickly exposed in the secondary market. As of March 2026, Zhipu's core MaaS (Model-as-a-Service) API platform had an ARR of about 1.7 billion RMB. However, its peer Anthropic's ARR was about $9 billion at the end of 2025, surpassing $30 billion within a few months—a gap not in the same order of magnitude. (Source: The Paper, May 8, 2026)
Going public solved the liquidity problem, but not the fundamental problem.
Zhipu HK market cap (May 2026): ~434.7 billion HKD, up ~16x from pre-IPO valuation (Source: The Paper)
MiniMax HK market cap (May 2026): ~257.3 billion HKD, up nearly 4x from listing (Source: The Paper)
Moonshot AI private market valuation: Surpassed $20 billion, cumulative financing over 37.6 billion RMB (Source: LatePost)
DeepSeek valuation (private market negotiation price): ~$45 billion, led by the National Integrated Circuit Industry Investment Fund (Source: 36Kr)
Third: Pricing by National Strategy
In the first half of 2026, capital's pricing logic for large model companies made its second leap. This switch was faster and larger in scale.
The National Integrated Circuit Industry Investment Fund led the investment in DeepSeek—the first time it directly invested in a major AI model company. The signal was clear: given potential compute constraints, China needs independent model companies that do not rely on NVIDIA or US cloud services. Large models were elevated to the same strategic level as chip manufacturing. (Source: 36Kr, May 8, 2026)
Meanwhile, Silicon Valley was simultaneously raising the reference ceiling: OpenAI's post-money valuation reached $850 billion, and Anthropic was raising a new round with a target valuation of $900 billion. Against this reference, the combined valuation of the Four Dragons—just over 1 trillion RMB, roughly $140 billion—was less than one-sixth of OpenAI's valuation alone.
This dual boost of 'strategic value + international reference' made it possible for three major model companies to raise over $10 billion in combined financing within three days. (Source: 36Kr, May 18, 2026)
Third time: Capital priced by 'national strategic score,' but whether this score is backed by fundamental business performance remains a question mark.
Putting These Three Upendings into a Framework
I call this pattern the 'AI Track Pricing Leap Model.' Each leap in pricing logic follows the same structure:
The old pricing logic is broken by a technological/strategic event (DeepSeek R1 / HK listing / National fund entry).
A new pricing logic is quickly established, with the primary market rushing to lock it in, followed by the secondary market.
The flaws of the new pricing logic gradually become exposed in fundamental business performance data.
Await the next technological/strategic event to trigger the next leap.
This framework has two important implications:
First, each leap is real, not a scam. The decline of the technical barrier is real, the HK window is real, the national strategic backing is real—these are variables with substantive content. It's just that each time, capital prices in the 'final state,' not the 'intermediate state.'
Second, the real watershed is ARR. The valuation of large model companies ultimately depends on Annual Recurring Revenue. Anthropic's ARR skyrocketing from $9 billion to over $30 billion in a few months is the foundation for its $30 billion+ valuation level. Among China's Four Dragons, the largest ARR is less than one-thirtieth of Anthropic's.
So This Round, It's the 'Five Giants'' Turn
The term 'Four Dragons' originally referred to Zhipu, MiniMax, Moonshot AI, and Baichuan. DeepSeek was occasionally included early on but remained on the periphery due to its non-commercial route. Now, the terminology is shifting to 'Five Giants,' because DeepSeek has received formal endorsement from the National Fund, making its valuation anchoring highly significant.
The shift from 'Four Dragons' to 'Five Giants' is not just a headcount change. It signifies that after the third leap in pricing logic, the ranking method in this sector has also changed: no longer based solely on 'model capability score,' but now with a 'national strategic weight' added.
Where will this new pricing system be challenged next?
On ARR. Once the story of national strategic backing is told, capital's eyes will turn to each quarter's ARR growth rate. Then, those with more solid fundamentals will defend their valuations; those buoyed by the pricing logic leap will face mean reversion.
The fourth pricing leap for China's large models is already on the way.
This article is for informational sharing and industry analysis only, and does not constitute any investment advice, investment analysis opinion, or trading invitation. The market carries risks; investing requires caution. Any investment decisions made based on the content of this article entail risks and profits/losses borne by the individual. The author and publishing platform assume no legal responsibility.
