At two in the morning the robotic bench is still working. Arms lift trays of vials, weigh reagents to the microgram, and load them into racks; a screen above the bench shows which of several hundred parallel reactions is behind schedule.
Nobody in the room is deciding what to try next. The plan was made earlier, in software, by a model that proposed candidates, ranked them and wrote the protocol; the machines are executing it and logging every deviation.
This is what China now calls an autonomous laboratory (自主实验室), and it is the unit its science policy is buying. The pitch is not a smarter model. It is a cheaper loop between a guess and its test.
Key takeaways
- Sichuan launched the Tianfu Huiyan (天府·慧研) science platform on 22 October 2026 around a model with about 35 billion active parameters (China News Service, 23 October 2026).
- Its operator reports a 20% rise in task completion and about 70% lower compute use — published by the operator itself, not independently audited (via China News Service, 23 October 2026).
- The platform's domestic-accelerator cluster runs at 408 petaflops, with a target of 600 petaflops by the end of 2026 (Chengdu Daily, 24 October 2026).
- Beijing's AI-for-science plan, published on 1 July 2026, sets 18 tasks and makes the autonomous laboratory the core delivery vehicle (Beijing Municipal Science and Technology Commission).
- The China National Institute of Standardization has issued a framework covering laboratory terminology, device interfaces and experiment data, CCTV reported on 24 October 2026.
What was switched on
Sichuan's platform, launched at the national computer congress in Chengdu on 22 October, is organised as one science model plus three supports plus a set of applications — in the province's phrasing, "1+3+N". The supports are domestic compute, curated scientific data and dry-and-wet laboratories.
The numbers attached to it are specific. The science model has about 35 billion active parameters and was post-trained on roughly 2 million papers from materials and health research; the data layer opens with a 200-terabyte secure data space (安全数据空间), and materials alone has pooled about 1 million records and 1 million papers (Chengdu Daily, 24 October 2026).
One agent answers questions over a database of more than 150,000 inorganic compounds; literature search sits alongside it. The whole thing was named on 20 October to a first batch of signature application scenarios selected by the National Development and Reform Commission and other departments (China News Service, 23 October 2026).
Why the money goes into plumbing
Beijing's plan, published on 1 July 2026, is the clearest statement of intent. It sets 18 tasks across five directions, names six domains for flagship scenarios — high-energy physics, materials, health, life sciences, quantum technology and biological breeding — and designates the autonomous laboratory as the vehicle rather than the model (Beijing Municipal Science and Technology Commission, 1 July 2026).
The money follows the same logic. The plan names municipal fixed-asset investment, capital science and technology vouchers (科技创新券) and local-government special bonds, and promises a construction standard plus tiered evaluation for autonomous laboratories — instruments, rooms and integration, not a training run.
China's research system is unusually suited to this and unusually exposed by it. Beijing alone counts more than 90 universities, over a million research staff, 800-plus academy members and 145 national key laboratories, close to 30% of the national total (Beijing Municipal Science and Technology Commission, 3 August 2026). That is a lot of instruments behind a lot of incompatible software.
Autonomous laboratories are not a Chinese invention; the United States, Britain and Germany have run them in materials and drug discovery, and the standardisation gap is international (CCTV, 24 October 2026).
Verification is the constraint
A model that proposes experiments produces hypotheses faster than anyone can test them. That is not a triumph; it is a backlog. The published rationale for the Sichuan platform is exactly this: researchers complained that models were not specialised, compute was scarce, data was hard to find and verification was slow.
The design answer is a loop rather than a benchmark. The model proposes a candidate; the robotic bench synthesises and characterises it; the instrument data returns; the model updates. If the loop is real, the output is not a claim but a record — and a record can be checked by someone who was not there.
That is why the standards work matters more than the launch. The China National Institute of Standardization has released a framework covering laboratory terminology, device interfaces and experiment data, reported by CCTV on 24 October 2026, on the argument that instruments from different vendors do not speak to each other and integration is rebuilt every time.
What would make this citable
Three things are worth watching, none of which is a benchmark score.
- A peer-reviewed result produced inside a closed loop. A material synthesised by the robot and characterised by the same pipeline, with raw instrument files deposited, is evidence; a chart in a launch deck is not.
- Whether the standards are adopted outside the alliance. A framework drafted by the builders, for the builders, standardises one vendor ecosystem.
- Whether provincial platforms interoperate. Sichuan's data space is a provincial pool with its own schema; a second province building the same thing differently recreates the problem the standards are meant to solve.
Honest limitations
The 20% and 70% figures are the operator's own, with no published definition of "task completion rate", no baseline and no external audit. The 408-petaflop cluster is one provincial facility, not a national capability, and utilisation is not disclosed. The standards document is a framework and a route map, not an approved standard. No peer-reviewed discovery has been attributed to the platform yet, and the six flagship domains in Beijing's plan are intentions whose budgets remain unallocated.
What readers can do now
- Before citing the 20% and 70% figures, ask what "task completion" measures and which baseline it is measured against.
- Check whether a paper claiming an AI-discovered material reports synthesis in the same closed loop, and whether the raw instrument data is deposited.
- For procurement, price the integration and the standards compliance, not the model licence; that is where the multi-year cost sits.