A venture partner in Shanghai closed her laptop halfway through a pitch. The deck described yet another large model (大模型); she had sat through more than forty like it that quarter, and the math no longer worked. A year earlier, that same slide would have earned a meeting. Now it earned a polite exit.
That quiet closing of the laptop is the whole story of China's AI primary market in 2025: the model bet is cooling, and capital is asking harder questions.
The numbers that changed the mood
- PitchBook's vertical tally: in 2025, VC money into China's AI sector totaled about US$6 billion across 574 deals — a significant drop from 2024's annual total.
- The comparison abroad: US AI deal value reached US$174.6 billion, up 61.5% versus 2024; Europe reached US$21.9 billion, up 19%. China's private market cooled even as the West accelerated.
- The state backstop: in March 2025 Beijing launched a 1 trillion yuan (≈ US$140 billion) government-backed fund for emerging tech including AI, semiconductors and quantum computing — a strategic floor, not a substitute for commercial returns.
Why the model bet stalled
Several forces pulled at once. The model-building layer saw fewer and smaller rounds, as investors stopped funding "another large model (大模型)" on promise alone. China's biggest internet firms — Alibaba, Tencent, ByteDance — pulled capability in-house, shrinking the field of external bets they once made. And public markets began demanding evidence of revenue, not parameter counts.
The crowding made it worse. When a dozen well-funded teams chase the same foundation-model dream, only a handful can remain standing, and investors know it. Capital concentrated in the names already considered safe, and the long tail of "also-rans" found the room empty.
Where the money went instead
The flow did not disappear; it redirected:
- Toward applications, vertical AI, compute infrastructure and embodied intelligence (具身智能) — places where a customer already exists.
- Toward proven commercial traction. Two bellwethers reached Hong Kong: Zhipu AI (智谱) listed on the HKEX on 8 January 2026, raising about HK$4.3 billion (≈ US$550 million); MiniMax listed 9 January 2026, priced at the top of its range to raise about US$538 million at a roughly US$6.5 billion valuation.
- Toward private strength over a rushed listing. Moonshot AI (月之暗面), maker of the Kimi assistant, closed a US$500 million C round in late 2025 rather than hurrying to market.
The revenue reckoning
Investors now ask three colder questions: real orders? Recurring revenue? Scalable delivery? The Hong Kong listings rewarded companies that could show commercial footprints — MiniMax with a large consumer user base, Zhipu with enterprise deployment — over those still optimizing benchmark leaderboards.
The public market told the same story from the other side. PitchBook noted that China's AI core conglomerates posted the strongest year-to-date returns globally, with Alibaba up 113% — a sign that listed strategic winners were bid up even as private rounds shrank. The enthusiasm is real; it is just narrow.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the market has moved from rewarding model size to rewarding a defensible revenue line, and that shift is now irreversible.
What it means for founders and LPs
- The "spray and pray" era of model bets is closing; round sizes concentrate in fewer, later, revenue-backed deals.
- The 1 trillion yuan state fund absorbs strategic risk, but private capital still prices the upside — and private capital is the one that exits.
- Geographic gravity is shifting. Shanghai's share of large-model deal value rose sharply in 2025 even as Beijing's fell, a reminder that talent and customers, not just capital, decide where the next winner forms.
Honest limitations
- The US$6 billion / 574 deals figure is PitchBook's vertical tally for 2025; methodologies differ across data providers, and China's true private flow — including undisclosed rounds and foreign participation — is partly opaque.
- Zhipu's and MiniMax's listing sizes and valuations are from Reuters and established financial-media reports at the time of pricing; market capitalizations moved on debut and afterward.
- We deliberately omitted lower-quality domestic aggregator estimates of model-layer funding declines because they lacked an authoritative primary source; we kept the trend qualitative rather than invent a precise percentage.
- Not investment advice; we hold no position in any named company.
What readers can do now
- Watch HKEX filings and prospectuses: the Zhipu and MiniMax documents are the first audited look at Chinese model-company revenue mixes — read the revenue breakdown, not the headline.
- For founders raising now, lead with a real customer and a recurring-revenue line; a benchmark screenshot no longer opens doors.
- For LPs, separate the 1 trillion yuan state fund (strategic, patient) from commercial VC (return-seeking); they play different games and should not be conflated.