The office block in Xiong'an is new, the nameplate is newer, and behind the glass there is not much to see yet: no product line, no customer list, no published price.
What exists is a mandate, written into a company registration and announced in one paragraph by the asset regulator. For a decade the binding constraint on Chinese artificial intelligence has been described in one word — chips — and the answer has been factories.
Ask the people who assemble training sets and a second answer appears, less photogenic: records they are allowed to use, cleaned to a standard, with provenance that survives an audit. Scattered files inside ministries and state-owned plants are not that.
The State Data Group (国家数据集团) is China's attempt to build the middle layer: an institution that takes raw administrative and industrial records, turns them into something a model can be trained on, and sells the result.
Key takeaways
- China's asset regulator announced on 20 October 2026 that the State Data Group would sit in Xiong'an, ranked 28th among 100 central enterprises (SASAC).
- The National Data Administration reported 52.26 zettabytes of data produced in China in 2025, up 27.28% year on year and about 27.44% of the world total (NDA, 29 April 2026).
- China counted more than 116,000 high-quality datasets (高质量数据集) holding over 960 petabytes as of end-March 2026 (National Data Administration, 29 April 2026).
- Of 2.53 zettabytes stored in 2025, only 0.56 zettabytes was structured — the form that can be queried and trained on (NDA, 29 April 2026).
- The group's stated task is to supply datasets and corpora (语料库) to the AI industry (Xinhua, via SASAC, 21 October 2026); it has claimed no ownership of platform behavioural data.
A company, not a bureau
The announcement was one paragraph long. On 20 October the State-owned Assets Supervision and Administration Commission said the State Council had approved the creation of the State Data Group; the company was unveiled in Xiong'an the same day, entering the central-enterprise list at number 28 and bringing SASAC's total to 100 (SASAC, 20 October 2026).
A company rather than a commission is a deliberate choice. A commission allocates; a company transacts. The shareholder list says what kind of company it is: the State Council, the Xiong'an New Area administration, the three telecom carriers and the two grid operators all hold equity (China Securities Journal, 21 October 2026). The entities that own the pipes and the customer records are the entities that own the wholesaler.
Xinhua, citing the company, described three roles: a national team for data-factor services (数据要素), a leader in interconnecting data infrastructure, and a strategic force in data technology. The sentence that matters for AI is shorter: the group will provide high-quality datasets and corpora to the AI industry (Xinhua, via SASAC, 21 October 2026).
Where the data actually is
Volume is not the problem. The National Data Administration's third annual survey put national data production at 52.26 zettabytes in 2025, a 27.28% increase and about 27.44% of the world total; roughly nine-tenths of the increment came from enterprises (NDA, 29 April 2026).
The shape of the stock is the problem. Three ratios from the same survey explain why a dataset sold to a model lab stays scarce where data is abundant:
- 52.26 zettabytes were produced, but 2.53 zettabytes were stored, up 21.05% (NDA, 29 April 2026).
- Structured data reached 0.56 zettabytes, or 22.13% of the stored total, while growing 43.59% — the fastest-moving slice (NDA).
- Public data released for authorised operation (授权运营) rose 53.96% in 2025, faster than publicly opened data at 31.71% (NDA).
Read together: the country produces enormous volumes, stores a small fraction and keeps most of what remains in forms that cannot be queried. The scarce object is not bytes but a labelled, rights-cleared, documented slice small enough to fit on a disk.
What a wholesaler can fix
Three failures are structural.
- Format. Ministries, provinces and state-owned plants describe the same object differently; a national layer can impose one schema.
- Provenance. A buyer cannot audit where a record came from; a wholesaler that holds the chain can issue terms.
- Licensing. Public data is authorised for use through provincial and ministerial operators; a central counterparty can consolidate the contracts.
What it cannot fix is ownership. The most useful behavioural data for training sits inside Tencent, ByteDance, Alibaba and their logs; a state wholesaler has no claim on it and has asserted none. Nor can it fix price: no price list has been published, and China has no established market price for a high-quality dataset.
The second constraint is jurisdictional. Data leaving China is governed by the 2024 cross-border rules and by the "important data" (重要数据) designation, which is decided sector by sector. A state wholesaler does not resolve that; it inherits it.
Three tests worth watching
Three disclosures would show whether this is a broker or a filing cabinet.
- Delegation. Whether ministries and provinces hand their authorised-operation (授权运营) contracts to the group or keep running them locally. The first transfer agreement is the signal.
- Pricing. Whether a public price appears for a named dataset, per licence or per gigabyte. Until then, the group is a cost centre with a mandate.
- Foreign access. Whether a non-Chinese buyer can license a dataset for training outside China, and on what terms. In 2025, 142.34 exabytes crossed the border, up 14.88% (National Data Administration, 29 April 2026).
Honest limitations
The group was unveiled on 20 October 2026 and has published no financial statements, headcount, asset list or prices; everything above about its role comes from government statements of intent. The 52.26-zettabyte figure measures data generated, not data usable for training. The 116,000-dataset count carries no public definition of quality, no licence list and no usage statistics, and whether the group will own data, broker it or merely set interface standards is unresolved.
What readers can do now
- Ask any dataset vendor which ministry or province authorised the underlying data, and request the licence terms rather than a compliance certificate.
- Watch SASAC and National Data Administration publications for the first authorised-operation transfer to the group; no transfer a year from now is itself evidence.
- For procurement, test whether the same dataset can be processed offshore. If it cannot, budget onshore compute before pricing the deal.