On 16 October, China's State Council executive meeting, chaired by Premier Li Qiang, heard a progress report on the "AI+" initiative. The readout published by Xinhua gives the subject one dense paragraph before the meeting moves on to the Teachers Law, the Price Law and a batch of administrative regulations. There is no new fund in it, no target number, no list of favoured technologies. That absence is the point. A cabinet-level review is the stage at which a campaign stops being a slogan and starts being a budget line.
Deployment pulls the technology
The first instruction is that policy should "use deployment applications to pull technological innovation forward", starting from what the country actually has: a complete industrial system, rich application scenarios and strong engineering capability. That is a deliberate inversion of the model most AI industrial policy assumes, in which governments fund research and then hope someone finds users. Here the user is supposed to come first, and the technical problems are supposed to surface from use.
This changes what the state is for. Its leverage shifts from grants to procurement and standards. The measure of the initiative is no longer how many models are released or how they rank, but whether the systems are still running in a factory, a grid dispatch centre or a port six months after the demo. The readout does not say that explicitly. Its emphasis on "converting into real productive forces" carries the same meaning.
The base is physical
The second cluster is the most concrete sentence in the document: strengthen technical research and algorithm innovation, "promote compute-electricity coordination", and "increase the supply of high-quality data".
Electricity appearing inside an AI policy sentence is not decoration. Large training and inference clusters are constrained less by accelerator supply than by where the power is, when it is available, what it costs. Compute-electricity coordination implies siting decisions, grid connection rules, load scheduling and probably tariff design: the unglamorous machinery that determines whether expensive clusters run at 40 per cent utilisation or 85 per cent. If this line turns into published tariff and grid-access rules, it will move more money than any grant programme announced this year. If it stays a phrase in readouts, it will not — and that is the easiest thing to check later.
Data gets equal billing, with "high-quality" doing the work. China does not lack volume; it lacks data that is documented, licensed and clean enough to train on. That points at the boring industries underneath AI — annotation standards, provenance, and the question of who pays to clean data that firms regard as their competitive advantage. "Increase the supply" is an instruction to someone, but the readout does not say to whom.
Pilot validation is the bottleneck
On applications, the meeting asks for the real economy as the focus, better matching of supply and demand, joint research between users and developers, progress in "pilot-scale validation", and more cultivated scenarios.
Pilot validation is the phrase that separates industrial AI from consumer AI. The gap between a model that works on a demonstration line and one that survives a real shift is not mainly a modelling problem. Tolerances drift as tools wear. Sensors get dirty. The shift supervisor does not trust a recommendation he cannot explain, and the plant manager will not risk a stoppage to prove a vendor's benchmark. Closing that gap takes testbeds, instrumented lines, calibration data and engineers who are allowed to fail in production — expensive, unglamorous, and usually shared, because no single factory wants to fund a testbed for a product it does not own. Whether joint research between users and developers fixes that depends on procurement rules that let large buyers share the cost and risk of a line trial — an institutional problem, and where most industrial AI programmes quietly stall.
Rules by sector, in tiers
The third instruction is the least quantified and the most consequential for anyone selling into China: coordinate development and security, "apply policy by industry, tier and category", strengthen risk assessment, and hold the safety bottom line.
Tiering means the compliance burden will not be uniform. A customer-service assistant, a medical triage system and a model touching industrial control ought not to face the same approval path, and the readout implies that they will not. But it names no sectors, no tiers, no regulator and no timetable. Until those catalogues exist, companies face the familiar interim problem of guessing which tier they are in — and the cost of guessing wrong is a launch that stalls in review. The direction of travel is nonetheless clear: safety is being designed as a sector-specific regime rather than one horizontal law, which is broadly what firms with regulated customers have asked for.
The same meeting priced the discounts
The AI paragraph is not the only relevant item. The same meeting discussed a draft amendment to the Price Law and used language that anyone selling AI capacity should read twice: govern "disorderly low-price competition" in accordance with the law, improve supporting rules for industry cost investigation and for stopping below-cost dumping, and keep regulating price competition in the platform economy.
Read beside the AI+ review, the pairing looks deliberate. The cabinet wants AI deployed at industrial scale, and it does not want that scale bought with losses. Chinese AI has spent two years in a price war, with inference prices cut repeatedly and free tokens used as marketing. A cost-investigation mechanism gives regulators a factual test — is the price below audited cost? — where previously they had a complaint and little else. Whether enforcement reaches model APIs is an open question; the machinery, once built, would not be limited to consumer platforms.
What to watch
Four checks, in rough order of how soon they should become visible. First, whether compute-electricity coordination produces published rules on tariffs, grid access or siting. Second, whether pilot validation acquires named facilities, standards and a funding owner. Third, whether tiering produces catalogues listing sectors, tiers and regulators, since that is the document companies will plan against. Fourth, whether price enforcement touches cloud and inference pricing — the first hard evidence that the state means to discipline AI competition rather than merely encourage it.
The phrase "cultivating scenarios" will be read by provincial governments as a key performance indicator, and the predictable output of a scenario KPI is a wave of demonstration centres built to be visited. Readouts cannot enforce themselves.
The harder test
The most consequential phrase may be the plainest: AI is "a key variable affecting future development". Variables are judged by what they change. For two years the public scoreboard for Chinese AI has been model launches, benchmark tables and the size of training runs, all of which are inputs. A cabinet-level review that asks about deployment, power, data and pilot lines is a statement that the scoreboard is changing.
It is a harder test for the companies that lobbied for the initiative: easy to demonstrate a model, much harder to show a plant running better because of one, with numbers an auditor would accept.
Sources
State Council executive meeting readout, "Li Qiang chairs a State Council executive meeting; briefing on the progress of the AI+ initiative heard", gov.cn / Xinhua, 16 October 2026 — https://www.gov.cn/yaowen/liebiao/202610/content_7083431.htm