The instruction that left Beijing at the end of October was not an instruction to build anything. Officials from the data regulator sat with counterparts from the pilot provinces and cities, went through progress, and asked them to work backwards from demand instead of forwards from capacity.
For two years the programme had run the other way. Collect, annotate at scale, then look for a buyer.
The reversal changes which published numbers mean anything.
Key takeaways
- China's data regulator describes a national annotation network of seven base cities plus 32 pilots, and says the first group annotated more than 142 petabytes (National Data Administration, 25 October 2026).
- That work served 482 large-model development projects, the regulator says — a supply figure, not a revenue figure (National Data Administration, October 2026).
- Nationally, more than 126,000 high-quality datasets and 1,815 petabytes existed by August 2026, up 89% from March (National Data Administration, October 2026).
- About 1,770 datasets had been published on the national registry by October 2026, the system the ministry now wants builders to use (National Data Administration).
- A four-ministry plan issued on 13 January 2025 targets compound annual growth above 20% for the annotation industry through 2027 (NDRC and three others).
What the network actually is
China's annotation programme is organised as a set of cities, not a set of companies. The National Data Administration (国家数据局) describes a network it calls "7+32": seven cities that took the first annotation-base task in early 2025 — among them Chengdu, Shenyang, Hefei and Changsha — and 32 more that began pilot work later (National Data Administration, 25 October 2026).
The vocabulary appears in every procurement document. A dataset (数据集) is the product; annotation (数据标注) is the labour; a chain-lead unit (链主单位) is the organisation the ministry designates to organise one sector's data supply.
Scale first, structure second. On 28 October 2026 the administration met 21 chain-lead units drawn from nine fields: education, health, research, industrial manufacturing, bio-manufacturing, agriculture, energy, transport and meteorology. A day earlier it met data officials from the 39 pilot provinces and cities. Both readouts were published on 31 October 2026.
The ministry changed the instruction
The phrases from those two meetings are administrative, and they are the story. "Supply set by demand" (以需定供) becomes a dataset demand map, compiled by surveying the firms upstream and downstream of each pilot city. "Build through use" (以用促建) makes deployment the test of a dataset's worth. And the registry is the enforcement point: annotated datasets are to be logged in the national system, quality-tested, and then tracked for reuse.
Read plainly, this is a correction. A programme that had been measuring itself in petabytes concluded that volume was not the constraint it had assumed.
The ratio to watch
Two sets of figures, published months apart, sit oddly beside each other.
- By June 2026, China had more than 120,000 high-quality datasets totalling over 1,565 petabytes, up more than 60% from the end of the first quarter. Seventy-two chain-lead units accounted for over 45 petabytes across more than 759 high-value scenarios; 140 pilot units accounted for over 106 petabytes (National Data Administration figures reported by Shanghai Securities News, 28 October 2026).
- By August 2026 national totals had reached more than 126,000 datasets and 1,815 petabytes, 89% above March (National Data Administration, 25 October 2026).
On those figures, the two organised groups hold a little under a tenth of the national total. Most of the petabytes sit with units outside the designated programmes: the state's map covers a minority of the supply it is trying to coordinate.
Then the registry. By October 2026 the national system had more than 770 registered entities and about 1,770 published datasets. Against 126,000 built, that is a very small share, and the honest reading is directional rather than exact: registration is new, voluntary, and the ministry has only begun to push it.
Who pays
The 2025 opinion that shaped the industry is explicit that annotation services should enter government procurement, and that local voucher schemes — data, algorithm and compute vouchers — can be used to cut annotators' costs. The same document sets the growth target and names 13 policy measures (NDRC, National Data Administration, Ministry of Finance and Ministry of Human Resources and Social Security, 13 January 2025).
So the money is municipal before it is commercial. A pilot city funds a base, staffs part of it through vocational colleges, and counts the payroll. The buyer is meant to be a chain-lead unit or a model developer; the payer, initially, is often the city.
The result is capacity that exists because a budget existed. The October instruction tries to attach that capacity to orders before the next building round.
What it means from outside
First, this supply is not purchasable. The datasets sit behind a domestic registry and, in many cases, behind the data-export rules this column has described. A foreign lab cannot buy its way into those 142 petabytes.
Second, the volume claims will keep arriving. When a partner or a ministry cites petabytes, the useful question is not how much was built but how much is registered, tested and reused.
Third, the quality story is changing shape. The vocabulary in October was about expert networks rather than headcount — annotation treated as a knowledge job, priced by specialism.
Honest limitations
Every petabyte and dataset count here is reported by the units that built the data, not independently audited. The registry share compares two different systems at two different dates, so it shows direction rather than a precise ratio. No unit-level pricing, utilisation or margin data is published, which means the economics of an annotation company cannot be read from these figures at all.
What readers can do now
- Check the national registry at ndsms.cn, count the quality-tested entries, and repeat the count next quarter. Growth in registered and reused datasets is the only public evidence that the demand-first turn is working.
- When a Chinese counterpart cites dataset volume, ask for the registration number and one reuse record. Without those two things, the number describes inventory rather than capability.
- If you are tracking market entry, follow procurement notices in the 32 newer pilot cities rather than the seven original bases. The 2027 growth target has to come from the new group.