At a State Council Information Office press conference on 19 October on technology-enabled commerce, the Ministry of Industry and Information Technology's science and technology director, Wei Wei, gave the ministry's accounting of the 'AI + Manufacturing' action plan it issued late last year with six other departments. According to MIIT's own estimate, China's core AI industry passed 1.2 trillion yuan in 2025.
The plan has covered 20 industries including steel, autos and pharmaceuticals, with more than 1,000 high-value application scenarios catalogued, 285 typical cases published, and supply-demand matchmaking for over 5,000 companies; MIIT says AI is now applied by more than 30% of above-scale manufacturers. Under a companion effort on data and models, the ministry reports more than 500 high-quality datasets across 20 industrial fields, over 100 industrial agents promoted, roughly 100 national and industry standards drafted, 70-plus international standards led, and an agent interconnection protocol released for uptake. A national standard series for grading the intelligence of AI terminals has been issued for seven product categories, including phones, PCs, TVs and glasses; MIIT says more than 100 products now meet the 'assistive' level and that shipments of smart terminals passed 100 million units in the first half of 2026. By end-June the country counted more than 6,600 AI companies, with an open-source community (AtomGit), a 60 billion yuan national AI industry investment fund and ethics review rules in place.
On humanoid robots, MIIT says more than 400 whole-machine models appeared in the first half of the year with over 40,000 units produced, and expects more than 100,000 for the full year, with trials in 300-plus scenarios such as warehousing, retail and home services. Next steps it named: a 2026 edition of a national humanoid-robot standards system, standards for scenario deployment, an open-source robot community, and product, network and data-security testing.
Our takeThe numbers that matter are about interfaces, not the trillion-yuan headline: a published agent-interconnection protocol and a seven-category grading standard for terminal intelligence try to make 'how capable' a property that can be tested and compared, rather than a demo claim. The 30% adoption figure is the weakest, because no definition of 'applied' is given — a pilot and a production line count alike. Watch whether the grading standard ships test methods with its labels.