On November 17, the state-run China Industry News ran an analysis built around a newly fashionable job title: FDE, or forward-deployed engineer.
Behind the title is a document. The general office of the Ministry of Industry and Information Technology issued a notice on a special programme to cultivate AI application service providers. Its reference number is Gongxin Ting Ke Han [2026] No. 414.
Two steps, 1,000 firms a year
The targets are concrete. By the end of 2026, the national pool of qualified service providers should exceed 2,000 companies. By the end of 2027, it should be no fewer than 3,000.
That means a net gain of 1,000 in a single year — about four new registrations every working day, assuming 250 working days.
The notice goes to provincial industry authorities, 32 reporting units in all. Spread evenly, 2,000 firms works out to a little over 60 per province. Provinces that host a national pilot zone for AI industrial applications must reach at least 100 local firms by the end of 2027.
Task one is specific: each region surveys local providers and builds a file on them, and the ministry aggregates the results into a national pool. The ledger comes first; grading comes later.
A service provider is not a systems integrator
The notice defines the category as companies that supply consulting, delivery, operations and security governance around a customer's automation needs.
The operative word is delivery. For two years, enterprises bought model tokens. Now they are buying the promise that a system runs inside their own plant.
China Industry News quoted several experts. Lu Feng, head of the Beijing Frontier Future Technology Industry Development Research Institute, described providers as the last mile between technology supply and industry demand, there to solve the problem of firms that cannot use, cannot afford, or dare not use AI. Ba Shiguang, deputy secretary-general of an embodied AI data alliance in Beijing's Shijingshan district, called the pool an industrial ledger, a matching base and a capability grading system.
What an FDE actually does
One line in the notice is easy to skip: providers are encouraged to build teams of forward-deployed engineers who sit on the customer's site.
The role has existed in Silicon Valley for several years. The engineer moves into the client's office and rewires the system line by line, because the client's business process does not match the product's assumptions.
That describes a market where standard products no longer sell. The same model behaves differently in a mine and in a hospital, and if nobody does the adaptation in between, the project stalls before acceptance.
The notice also opens two purchasing channels: first-purchase programmes and risk compensation. What can be bought is defined broadly — models, agents, and token-billed services.
The hardest of the four tasks
The document lists four tasks: build the pool, improve supply, drive scaled adoption, and strengthen support.
The third is the most concrete. Providers are told to package modular, standardised solution bundles around high-frequency, rigid-demand, reusable business processes — small, fast, light and precise products.
In plain terms: stop selling custom projects. A custom project bills once and dies when the team leaves. A product bundle loses money in research and earns it back through replication. Most providers are stuck at that transition, which is why the 2,000-firm target needs a policy push.
After the ledger exists
Two thousand firms is not large against China's software and IT services industry. But it turns "who is actually helping companies use AI" into a list that can be named, graded and held accountable.
Companies on that list will compete on delivery cycles and repeat purchase rates, not parameter counts.
For buyers, the test is simpler: how many projects in my industry, how long did they take, and who runs the system after go-live.
What the pool changes for buyers
For a state-owned enterprise or a local government buying AI, the pool works as a pre-qualified list. Instead of evaluating dozens of unknown vendors, a buyer picks from a graded pool and inherits the ministry's screening.
That shifts the burden onto the vendor. Being in the pool is not a subsidy; it is a reference that can be checked, and removed.
The two-step design — 2,000 by the end of 2026, 3,000 by the end of 2027 — also implies the first year is about quantity and the second about quality. Firms that enter early and never deliver will be visible in the second count.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*