---
title: "China's Government Is Buying AI. The Rulebook Keeps It an Assistant."
date: 2026-11-13
category: China AI Decoded
site: NeuroAI
canonical: https://neuroai.site/a/na-decode-government-ai-assistive-rule
language: en
---

# China's Government Is Buying AI. The Rulebook Keeps It an Assistant.

> A national guideline fixes government AI as an assistant, pushes counties onto provincial platforms, and leaves a market that only private trackers count.

A county information official has a new and uncomfortable instruction. No model platform of his own, no separate contract for a system the provincial capital already runs. He is to connect to the provincial platform and use what it provides.

That instruction did not come from a budget circular. It is written into a national guideline for government large models, issued by the Cyberspace Administration of China and the National Development and Reform Commission. The document is more than a year old, but the machinery it points to — provincial platforms, a national standard — is only now being built, as provinces publish their own AI action plans.

The interesting question for anyone selling into this market is not how big it is. It is what the buyer is permitted to purchase.

## Key takeaways

- The government large-model guideline, published by the regulator on 10 October 2025, lists 13 reference scenarios in four categories and fixes AI in an "assistive" role.

- Public-facing services must run content review, using measures such as human checking, real-time controls and cross-model verification.

- Counties are told to reuse provincial compute and models rather than build their own.

- A national technical standard for government large-model systems, GB/T 48324-2026, takes effect on 1 April 2027.

- One private tracker counted 4,481 awarded large-model contracts in Q3 2026, with a median value of RMB 800,000.

## The rule, not the budget, sets the market

The guideline is unusually concrete: four categories — government services (政务服务), social governance, office work and decision support — and 13 reference scenarios inside them.

Then it constrains them. Public-facing scenarios such as question answering and assisted handling must run content review, drawing as appropriate on human checking, real-time risk controls and cross-model verification, and must carry a visible notice of the model's limits plus labels on its output. For common scenarios such as question answering and document drafting, departments must use mature products whose models have already completed filing (备案) with the regulator. The logic is the asymmetry of error: a wrong checklist wastes a trip; a wrong policy citation enters the record.

The most consequential clause is about who may build. Cities deploy under provincial rules; counties and below "should in principle reuse the compute (算力) and model resources of higher levels, and no longer independently build and deploy government large models." Provinces run one service platform and register regional compute, models and datasets in a single ledger.

As procurement policy, that shrinks the number of buyers. It also sketches a commercial model: a company builds and operates, government buys the service, billing follows usage.

## What the tender data can and cannot show

No official series measures this market, so two private trackers fill the gap. Their numbers are not comparable, because they count different things over different periods.

Intelligence Superparameters (智能超参数) counted 4,481 awarded large-model contracts in the third quarter of 2026, of which 3,500 disclosed a value totalling RMB 15.26 billion (≈ US$2.1bn / HK$16.8bn). The count was up 130% year on year and the disclosed value up 56.5%. Deals above RMB 10 million rose to 171, from 123 in the second quarter. The median contract, though, was RMB 800,000 (≈ US$113,000): growth is in the number of small purchases, not their size.

A second tracker, Qianlima, was searched by China Business Journal in October 2025 with the single keyword "government plus large model": 2,197 tender notices and 1,630 award notices for the first nine months of 2025, against 937 and 856 for all of 2024.

Both are samples. Neither shows whether the systems are used: a signed contract is a commitment, not an operating metric, and the guideline's own call to build an evaluation system suggests none yet exists.

## The standard that lands in April

The guideline said standards would follow. One has. GB/T 48324-2026, Artificial Intelligence — Technical Requirements for Large-Scale Model Systems of Government, was published on 28 September 2026 and takes effect on 1 April 2027.

Its drafting list is the most useful part: dozens of organisations are listed as drafters, including the China Electronics Standardization Institute, Huawei, the Beijing municipal big data centre, the e-government centre of the Fujian provincial government office, Dahan Software, Alibaba Cloud, iFlytek's government subsidiary and Baidu — ministries, platform operators and vendors in one document. That mix is how a technical requirement becomes a procurement checklist.

Two caveats matter. GB/T is a recommended standard, not a mandatory one; it binds through tenders that cite it, not through law. And it governs systems, not outcomes: a compliant platform can still return a wrong answer.

## What this means for a foreign vendor

Three things follow:

- The unit of sale is an assistant inside someone else's platform, not an autonomous agent with its own permissions.

- The customer of record is increasingly provincial, so a city-level relationship no longer converts into a contract alone.

- Compliance artefacts — filing numbers, evaluation results, labelling and audit logs — are part of the product, not paperwork around it.

Provinces are now writing the implementation layer. Fujian's 2026–2028 action plan, reported on 12 November 2026, promises more than 60 industry-specific models, more than 100 high-quality datasets, an expanded compute-voucher scheme and at least ten benchmark scenarios a year from each leading department. That is the shape of demand: many small, specified, local deployments.

## Honest limitations

There is no official statistic for China's government AI market; both counts cited here are private trackers using different keyword and scope rules, and neither publishes an audited method. The guideline is guidance, not regulation, and its language ("should", "in principle") leaves room for exceptions that no public document records. GB/T 48324-2026 is recommended, and its text was not reviewed here beyond the official registry entry. No vendor discloses government revenue separately, so share claims are estimates. Whether these systems improve service quality is not measurable from procurement data.

## What readers can do now

- Pull GB/T 48324-2026 from the national standards registry and read its scope before bidding: a system that fails the clause list will not survive a tender that cites it.

- Map your product against the guideline's 13 scenarios and note which are public-facing, because those carry human-review and labelling obligations.

- Track provincial platform tenders rather than city ones: the consolidation clause moves the buying decision up a level.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-government-ai-assistive-rule
Free to quote with attribution and a link to the original.
