China put industrial AI standards beside a public compute counter

THE SIGNAL IN ONE SENTENCE

A factory does not need another speech about transformation. It needs a machine that produces fewer bad parts on Tuesday afternoon. That is the useful tension inside the 2026 AI Industry Conference in Jinan. The China Academy of Information and Communications Technology, or CAICT, reported on October 16 that the conference had placed several pieces of industrial AI infrastructure beside one another. There was a map of more than 1,000 high-value scenarios across 20 manufacturing sub-industries. There was a 2026 safety evaluation framework for AI in industry and information technology. Five organizations began cooperating on a national AI application pilot base. A new public-service platform brought together computing power, models, data and applications. A separate meeting focused on connecting AI standards with manufacturing needs. There was even a new investment index intended to translate official AI industry classifications into something capital markets could identify, track and allocate money toward. Taken together, the announcements describe a machine for scaling machines. The plain signal is that the public now needs the operating receipt. A shared compute platform can lower the cost of trying industrial AI. A standards system can prevent every plant from inventing its own definitions. A scenario map can help factories find jobs worth automating. None of those instruments, however, proves that a smaller manufacturer can obtain affordable resources, integrate a model into old equipment, protect workers and data, or produce a measurable improvement after the conference banners come down. The next phase should connect each standard, compute allocation and factory deployment to field evidence. Start with the new China Computing Valley AI Public Service Platform. CAICT says the platform was jointly released by the Shandong Provincial Department of Industry and Information Technology, the Shandong Communications Administration, CAICT and the Jinan municipal government. The announcement describes it as part of China's computing-power platform system and says it aggregates four kinds of resources: compute, models, data and applications. Its stated job is to match regional computing resources and support vertical industry scenarios, while serving as an eastern hub in the national integrated computing network. That is the architecture in one paragraph. The factory question begins where the paragraph stops. Who can use it? What does access cost? How long does a job wait? Which hardware is available? Can a small manufacturer bring its own model and data? Can it take the trained artifact elsewhere? Who sees proprietary production records? What happens when a workload fails halfway through a run? A public platform should publish an access ledger that answers those questions without requiring every machine shop to become a cloud-contract specialist. For each class of user, the ledger should show eligibility, pricing, subsidies, queue time, service level and support route. For each resource pool, it should disclose hardware type, usable capacity, utilization, energy consumption and outage history. For each model or application service, it should state the licence, data-retention rule, portability format, security boundary and exit process. The point is not to drown a factory manager in infrastructure trivia. It is to make the public promise testable. If the platform is meant to match regional demand with computing resources, report the requests that were matched, the requests that were not, the time to allocation and the reasons for rejection. Separate large enterprises from small and medium manufacturers. Separate research trials from production workloads. Separate a one-hour demo from a system that runs a line every day. Access should be measured at the factory gate, not the platform launch. The conference's high-value scenario map offers a second piece of the puzzle. CAICT says the map covers more than 1,000 scenarios in 20 sub-industries and uses a three-level structure: research, validation and promotion. It organizes the material as one map for each industry and one file for each scenario. That is a sensible way to resist the great industrial temptation to call every camera and dashboard artificial intelligence. The stages should carry evidence requirements. A research-stage scenario can begin with a clearly stated production problem, a baseline and a technical hypothesis. A validation-stage scenario should add trials on representative equipment, failure cases, human review and comparison against the existing process. A promotion-stage scenario should require evidence from more than one site, known integration costs, maintenance needs and a record of incidents. Promotion should not mean that a vendor delivered a polished demonstration. It should mean that another factory can understand the conditions under which the result is likely to travel. Consider visual inspection. A model may detect surface defects on one product under controlled lighting. The useful field record includes the defect taxonomy, camera placement, line speed, false reject rate, missed defect rate and performance across shifts. It should show what happened when the material, supplier, dust level or lighting changed. It should record how workers handled borderline cases and how quickly the model was updated. The same discipline applies to predictive maintenance. A system that forecasts a bearing failure needs a comparison against scheduled maintenance, a cost for false alarms and a plan for sensors that drift. A useful result is not the number of alerts. It is avoided downtime after counting unnecessary interventions and missed failures. For process control, quality and safety cannot be separated. A model that increases throughput while narrowing a temperature margin may make the monthly chart look cheerful and the plant manager less so. Field tests should include operating limits, rollback conditions, sensor uncertainty and the person authorized to stop the automated recommendation. This is where standards become practical. The conference included work by the Ministry of Industry and Information Technology Artificial Intelligence Standardization Technical Committee, known as MIIT/TC1. A session scheduled for October 16 focused on connecting common technical standards with manufacturing scenarios. The announced participants covered industrial AI, metallurgy, building materials, machinery, home appliances, textiles, petrochemicals, shipbuilding and aluminium. That breadth is useful because industrial AI does not enter an empty software container. It enters furnaces, looms, assembly lines, inspection stations, warehouses and maintenance routines built under different physical constraints. A common standard should define the minimum evidence package while allowing each industry to add its own hazards. The common layer might cover data provenance, model version, performance baseline, monitoring, cybersecurity, human authority, update control and incident reporting. A textile application may add fabric variability and rapid product changes. A chemical plant may add process safety and strict operating envelopes. Shipbuilding may add long asset lives, distributed suppliers and inspection records that must remain legible years later. Interoperability belongs in the common layer too. Factories often inherit a pile of machines from different decades and vendors. A standard that works only inside one vendor's stack is less a standard than a velvet rope. Manufacturers should be able to export models, evaluation records, production data and event logs in documented formats. Interfaces should survive a supplier change. Security controls should follow the data instead of disappearing when it crosses a product boundary. The new Industrial and Information Technology AI Safety Evaluation Framework adds another important tool. CAICT says the 2026 framework covers general safety, application safety and business safety, and specifies evaluation content, methods and implementation requirements. That is a framework, not a clean bill of health for systems already deployed. The difference is easy to lose. A safety evaluation document can tell a team what to examine. It does not show which systems have been examined, by whom, against which version, with what result or under what factory conditions. Every deployment using the framework should produce a compact safety receipt. It should name the evaluator, system version, intended use, test environment, known limits, failed tests, mitigations, approval owner and next review date. If a vendor tests its own product, say so. If an independent laboratory repeats the work, say that too. General safety should cover the foundation: model integrity, access controls, data protection and resilience. Application safety should examine the actual industrial task and its failure modes. Business safety should ask whether the deployment can interrupt production, expose trade secrets, create legal risk or concentrate authority in a supplier that a factory cannot replace. Worker impact must sit beside those categories, not underneath them in small print. Industrial AI changes jobs even when no worker is removed. It may move inspection from seeing every item to reviewing exceptions. It may turn an experienced operator's judgment into training data. It may increase the number of lines one person monitors. It may make work safer by keeping people away from hazardous equipment, or more brittle by asking them to recover a process they no longer practice. The field ledger should record who gains discretion, who loses it and who carries the blame when the model is wrong. It should measure training time, override frequency, alert burden, near misses and whether workers can report a recurring problem without fighting the interface. A productivity claim without the labor denominator is a partial equation. Energy use deserves the same treatment. Shared compute can improve utilization by pooling expensive hardware. It can also hide the energy cost of training, inference and data movement behind a convenient endpoint. The public platform should report energy use per workload class and disclose when jobs are shifted across regions. Factories comparing models should be able to consider energy alongside accuracy, latency and price. The five-organization pilot-base collaboration offers a place to test these receipts. CAICT says Inspur Yunzhou Industrial Internet, Shanxi Jinyun Internet, Suzhou Digital Intelligence Technology Group, Shandong Port Qingdao Port Group and Haier Smart Home began cooperation on common AI support capabilities, high-quality datasets and shared application scenarios. The group spans platform, regional, port and manufacturing contexts. That variety can expose whether the standards work outside a single laboratory. The public record should show which components are shared, which datasets can actually be reused, what access restrictions apply and how performance changes across sites. One successful deployment at a large port or appliance manufacturer should not be treated as automatic proof for a smaller supplier. Scale brings engineering capacity, data volume and procurement leverage that many manufacturers do not have. Small firms need a different success test: time from application to first workload, out-of-pocket cost, hours of specialist support, integration with existing machines, staff training and the ability to leave without abandoning the work. Then there is the new CSI BOC Shanghai-Shenzhen-Hong Kong Artificial Intelligence Industry Index, compiled by BOC Fund, CAICT and China Securities Index. CAICT says it was designed to turn the AI standardization committee's industry classifications into a capital-market benchmark and that it received regulatory approval. An index can make a sector visible to investors. It cannot show whether a particular factory improved quality, cut energy use or reduced injuries. Market recognition and industrial performance may move together. They may also wave at each other from opposite sides of a spreadsheet. The index should be treated as a financial measurement product, not a proxy for factory outcomes. China's industrial AI programme now has several pieces that other regions are also trying to assemble: shared infrastructure, scenario catalogues, technical standards, safety evaluation, pilot sites and capital. The combination matters beyond China's borders because manufacturing supply chains cross jurisdictions. A model may be trained in one place, integrated into equipment in another and used to produce components sold everywhere. Comparable evidence can travel even when policies do not. A manufacturer in Germany, Japan, Mexico or Vietnam should be able to understand what a Chinese industrial AI system was tested to do, which standard it followed, what data it saw and where its guarantee ends. Chinese suppliers should expect the same clarity from foreign vendors. International recognition becomes more useful when it rests on inspectable test records rather than ceremonial alignment. The Jinan conference does not establish that this evidence system exists. Its announcements describe platforms, maps, frameworks, working groups and collaborations. They do not yet provide broad adoption data, audited productivity gains, equitable compute access, worker outcomes or factory-level incident records. That is the caution, and also the assignment. Publish one field ledger that links the pieces. Start with the scenario. Name the plant problem and baseline. Show the compute allocation and cost. Identify the model, data and standards. Record quality, safety, energy and labor results. List incidents, overrides and changes. State whether the system can move to another platform. The standard tells the factory what the instrument is supposed to measure. The public compute counter makes the instrument available. The field ledger shows whether anyone fixed the Tuesday problem. That is where industrial AI stops being a conference category and starts becoming a production result.