The fund manager keeps three decks in one folder, named after cities. On Monday he is in Guangzhou, walking a state-owned mother fund through a robotics portfolio. On Wednesday he is in Yibin, where the cheque is smaller and the decision is quicker. On Friday he is in a district of Wuxi that has money and wants a supply chain built beside it.
He is not raising from family offices. He is raising from municipalities, and the questions are not the usual ones. How many jobs, how much local procurement, when does the line start.
That is what Chinese AI venture capital looks like now.
Key takeaways
- Guangzhou committed RMB 20 billion (≈ US$2.8bn / HK$22bn) to an AI industry fund, built as a mother fund plus direct funds (Guangzhou municipal government, 14 April 2026).
- Direct funds of RMB 0.5–3 billion each (≈ US$70m–423m / HK$550m–3.3bn) carry a 4% hurdle; weak performers lose capital, strong ones can get three successive increases (Guangzhou municipal government).
- The city hosts more than 2,100 AI companies and 73 filed large models (大模型), third nationally (Guangzhou municipal government).
- City and district state-owned funds with an AI focus number nearly 200, holding more than RMB 160 billion (≈ US$22.5bn / HK$176bn) (Guangzhou municipal government).
- Yibin added RMB 2 billion (≈ US$282m / HK$2.2bn) and Wuxi's Xishan district set up RMB 3 billion (≈ US$423m / HK$3.3bn) (Ifeng; district government).
The structure is a tournament
Guangzhou's fund is not a single pool. It is a mother fund that allocates into direct funds, and the first three of those are RMB 0.5–3 billion each (≈ US$70m–423m / HK$550m–3.3bn) (Guangzhou municipal government, 14 April 2026). The mother fund sets the terms; the direct funds do the investing.
Each fund carries a 4% hurdle — the annual return it must clear before anyone shares the upside — and the city keeps score. Underperformers can lose capital; strong performers can be given three successive increases in allocation.
Officials call this "horse-racing" (赛马): the city is not choosing one strategy but funding several, and letting results decide where the next yuan goes. For a fund manager, a plausible thesis is not enough: someone else is pitching a rival one into the same review.
Why the scoreboard picks atoms
A municipal fund is judged on more than financial return. It is judged on what lands inside the municipal boundary: payroll, suppliers, factory floors, tax.
The consequence is structural. A software company can take an investment, hire in Shanghai and sell nationally, and the city that wrote the cheque captures very little. A robot integrator has to build a line, hire technicians, buy from local suppliers and train staff who cannot easily be moved. The second company is a better instrument of local policy even when it is a worse financial bet.
So the money tilts toward hardware, embodied intelligence (具身智能) and factory automation — categories where the asset is bolted to a postcode. Consumer applications are structurally underfunded relative to their share of usage.
Two smaller cities, same logic
Guangzhou is the loudest case, not the only one.
- Yibin, Sichuan: a RMB 2 billion (≈ US$282m / HK$2.2bn) AI fund on top of an existing RMB 9.5 billion (≈ US$1.34bn / HK$10.5bn) special fund, plus a RMB 2 billion list of application scenarios (Ifeng, 10 March 2026).
- Xishan district, Wuxi: a RMB 3 billion (≈ US$423m / HK$3.3bn) fund, with a first tranche of about RMB 1 billion (≈ US$141m / HK$1.1bn) (district government, 6 March 2026).
The pattern repeats at different sizes. A city with no chip industry funds robots; one with an industrial base funds automation for its own plants; one with neither funds a scenario list. The instrument is the same each time: public capital looking for a local footprint, and a fund manager paid to find companies willing to sign for one.
Four distortions worth naming
- Duplicate capacity. Every city that wants a humanoid cluster builds one, and the second and third plants are justified by the first, not by demand.
- Mispriced capital. A 4% hurdle is far below what a venture fund needs to justify its risk, which means deals clear that a market investor would reject.
- Relocation as a condition. A fund whose mandate is local output has every reason to tie capital to a local entity, plant or hire count, which taxes exactly the companies that are most mobile.
- Measurement by announcement. Funds report commitments, not deployed capital or returns, so the size of the programme is easier to know than its performance.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, these vehicles are better read as industrial policy with a return requirement attached than as financial capital hunting the best team. That framing explains the portfolio: the state is buying capability that stays put.
Who wins and who loses
The winners are hardware teams, robotics integrators and suppliers who can accept a local footprint; founders willing to move; and fund managers with a record inside the system. The losers are consumer-software founders who cannot use a factory, teams that refuse to relocate, and foreign co-investors who cannot work inside that mandate. For the companies that do win, the money is patient in a way that private capital is not. For everyone else, it is a reminder that the largest pool of AI risk capital in China is allocated by people whose job is regional development.
Honest limitations
Only Guangzhou discloses the hurdle rate and the tournament mechanics; Yibin and Xishan have published amounts but not full terms, so comparisons across cities are approximate. The "nearly 200 funds, more than RMB 160 billion" figure counts vehicles and commitments, not capital actually deployed, and no city publishes fund-level returns. The claim that this capital mix produces more hardware than software is an inference from incentive structure, not a measured outcome.
What readers can do now
- Operators: if you raise in China, treat the choice of city as a product decision, because the local footprint requirement will shape your cost base and your hiring for years.
- Investors: map the fund structures before mapping the companies. A cap table with three municipal vehicles behind it tells you what the company had to promise.
- Researchers: ask for deployed capital and fund-level returns rather than commitments; the gap between the two numbers is the story.