Executive Summary

On 20 October 2026, National Data Group Co., Ltd. (国家数据集团有限公司) opened in Xiong'an New Area (雄安新区), Hebei Province. It is a central state-owned enterprise in which the State-owned Assets Supervision and Administration Commission of the State Council, SASAC for short (国资委), acts as shareholder on the state's behalf. Chinese outlets reported that the firm entered the central SOE register (央企名录) at number 28, and that the count of central SOEs under SASAC supervision thereby returned to 100. Twenty-eight and 100 are different numbers. The first is a position on a list; the second is a total. Korean-language coverage in places turned this into "the 100th central SOE," a ranking that appears in no official notice.

What the company does was stated comparatively plainly by state media. Below is the core business as Xinhua's Korean-language edition put it, rendered here into English, with the original Chinese underneath.

The third of those three clauses is why this article takes up the company at all. Supplying datasets and corpora to the AI industry is now the declared business of a firm whose equity is entirely state capital. And the second clause sets up this article's question in advance. The same company also sells industry data governance services. The side that supplies and the side that appraises sit inside one legal entity, a thread picked up again in Section 5.

The fastest read on the company's character is not the business description but the shareholder list. Seven parties put money in, and five of them are in telecom and power. China Telecom (中国电信), China Unicom (中国联通) and China Mobile (中国移动), together with State Grid (国家电网) and China Southern Power Grid (中国南方电网), joined the State Council and the Xiong'an New Area administration as shareholders.

Why that list matters is simple. "Training data is starting to be treated as critical supply infrastructure, like power or telecom" is normally a line of interpretation. At this company it is the shareholder register before it is interpretation. The firms that build and run nationwide networks, the ones holding traffic records, metering records and facility logs, now divide the equity of a data supply company between them.

Plenty was not disclosed, though. The individual stakes of the seven shareholders were not published, and the notice carries no figure for registered capital. Ahead of the launch, Chinese financial outlets circulated estimates in the tens of billions of yuan, but that is a market guess, not a filing. The leadership is likewise known only through reporting. The chairman and party secretary is said to be Jian Qin (简勤), formerly general manager of China Unicom, and the general manager Chen Ronghui (陈荣辉), formerly a deputy head of the National Data Administration. This article could not open the original notice, so everything other than the business remit is written as "according to reports."

Putting the headquarters in Xiong'an is no accident either. Xinhua wrote that Xiong'an is the company's proving ground and incubator. Xiong'an is a new city designed to absorb functions moved out of Beijing, and it has been used as the place where new institutional arrangements are run first before being released nationwide.

Two organisations get mixed up because their names run close together: the National Data Administration (国家数据局) and the National Data Group. The first is a government body created in October 2023 that designs the underlying rules and policy. The second is a company created in October 2026 that actually gathers and sells data on top of those rules. Compressed into a line: the side that wrote the policy founded, two years later, the company that will carry it out.

The two are also linked through people. Chinese outlets reported that Chen Ronghui, general manager of the National Data Group, came from a deputy head post at the National Data Administration. Someone who sat where the rules were drafted now runs the company that operates under them. There is no basis for calling that improper in itself. It does, however, turn the question Section 5 will ask, who exactly plays the verifier separate from the supplier, into something other than an abstraction.

Another common misreading is "a company that appeared out of nowhere." State-owned firms that gather and sell data have been operating in China for several years already. A tally published by the Shanghai Data Research Institute in September 2025 counted 29 or more provincial-level data groups and 187 prefecture-level (地市级) ones, the latter amounting to 64% of China's prefecture-level cities. Their registered capital is not small either: Shanghai Data Group holds RMB 5 billion (about $700M) and Fujian Big Data Group RMB 10 billion (about $1.4B). The diagram below lays out those tiers.

Behind the three tiers sits a mechanism called authorised operation of public data (授权运营). The government entrusts public data it holds to a designated operating entity, which processes and circulates it. The National Data Administration said that as of 25 August 2026 the mechanism had begun operating in earnest in 25 of China's 31 provincial-level administrative divisions. Data groups appeared in the provinces first because someone had to be that operating entity.

Nor is this a mechanism in name only. A survey report from the same agency shows that the volume of data flowing through authorised operation rose from 4.93PB to 7.59PB in a year, up 53.96%. That is a record of the arrangement, setting up a local entity and entrusting public data to it, actually moving volume. The national-tier company was laid on top of that.

The timing deserves a look too. The "Data Elements ×" (数据要素×) three-year action plan, issued by the National Data Administration together with 17 other ministries, ran from 2024 to 2026, and the administration itself called 2026 the wrap-up year. The plan set two targets: to produce at least 300 exemplary application cases by the end of 2026, and to grow the data industry by more than 20% a year on average. At the point where three years of laying the rules come to an end, the company meant to carry those rules now exists at the national tier.

The picture the shareholder list showed is one piece of a larger design. On 9 June 2026 Bloomberg reported that China was preparing a plan to put roughly RMB 2 trillion, about $295 billion, into a nationwide data centre network over five years. It was described as a draft led by the National Development and Reform Commission, aimed at bringing a nationally unified computing network online by 2028, with a condition attached that more than 80% of compute chips be sourced from domestic suppliers. Ultra-long-term special treasury bonds and state investment funds were named as funding sources, and state-owned carriers such as China Mobile and China Telecom as operators. The same article carried a source estimate that the total could exceed RMB 5 trillion once grid reinforcement is included.

One thing has to be made plain here. That RMB 2 trillion is not money going into the National Data Group. The report predates the launch by four months, the parties named as executing it are different, and the figure appears nowhere in the coverage of the 20 October launch. Reading the two events as one current is possible, but that is interpretation, not an announcement. One word in the Bloomberg piece does support the reading, though. The data centre plan is introduced as part of a programme called the "six networks" (六网): computing, power, telecom, water supply, urban underground utility corridors and logistics, all laid as one bundle. Compute, electricity and water are on the same list.

One line for a sense of scale. The same Bloomberg article noted an expectation that US big tech would spend roughly $725 billion on AI in 2026 alone. China's five-year total is smaller than one year of American private spending. What this article is about is not the size of the money but the name on the column it flows into.

Whoever treats data as critical supply infrastructure starts counting it. China already tallies and publishes data volumes annually the way it does electricity consumption, and it publishes the tokens processed per day as something close to a national statistic. These are the most recent figures released by the National Data Administration.

Source: National Data Administration, National Data Resources Survey Report (2025), released 29 April 2026, and China Data Industry Development Report (2026), released 28 August 2026, via Xinhua coverage. Production total and storage total are different metrics and are better not read as one.

The top row of the table splits one layer further. The same report puts the share of that 52.26ZB produced by companies at 33.54ZB. As a proportion that is 64.18%, up from 58.30% the year before. The state is counting the total, but close to two thirds of that total comes out of companies. When the state says it will supply data, this ratio and the shareholder list in Section 1 point the same way about where that data comes from. That is also where the reason sits for putting the firms that build and run nationwide networks on the shareholder register.

One cell was left out of this table on purpose: the size of the data element market. The National Data Administration does not publish that item separately, and only private research estimates circulate. Those estimates sit anywhere from the RMB 100 billion range to the RMB 200 billion range depending on the house, a spread of more than two to one. The range shifts according to whether only data transactions are counted, or processing and services as well. Same country, same year, different yardstick, different number, which is the same grain as the question this article comes to later. A figure that does not say what it counted cannot be compared.

Korea answered the same question in its own way. A week before the Chinese launch, on 13 October 2026, the Cabinet approved amendments to the Act on the Development of Artificial Intelligence and Establishment of Trust, known as the AI Framework Act, and to the Framework Act on Promotion of the Data Industry. Within one week the two countries picked different vessels. China built a company; Korea built an obligation.

The core of the Korean amendment fits in a line. AI training data built or produced in the course of carrying out a government-funded project must be linked to and provided to AI Hub. The exemptions were left to presidential decree. Trade secret protection, infringement of personal data or copyright, and national security were the grounds named, with the precise scope to be set in the enforcement decree. That the project operator is the one who decides whether any of those four applies is something we read separately. On the data industry side, the amendment strengthens what is called a one-window system, where the location information for public and private data can be found through a single counter. The Minister of Science and ICT may ask public institutions and private entities to link their location information, and public institutions must comply absent special grounds.

The date is still ahead. The amendments take effect six months after promulgation, that is, in April 2027. Saying the obligation "is already in force" is not accurate as of now. And for this article, the subject of the quality clause in that amendment matters particularly. The parties required to take the measures necessary for stable quality management of linked and provided training data are the Minister of Science and ICT and the heads of the relevant central administrative agencies. Not the side that hands the data over, and not the side that uses it, but the side that receives and gathers it carries the quality duty too.

There was movement on the organisational side in Korea as well. On 1 October 2025, Statistics Korea was elevated into the National Data Agency (국가데이터처), 35 years after it opened in 1990. Headcount, though, rose by ten, from 689 to 699, and the newly created National Data Management Bureau holds 102 of them. Coverage at the time relayed an internal complaint that the division actually handling data management within it had four working-level staff. The table below places the two countries' designs in the same columns.

Source: for China, Xinhua, Chinese outlets reporting the SASAC notice, and National Data Administration releases. For Korea, coverage of the Ministry of Science and ICT press release of 13 October 2026 and reporting on the National Data Agency's launch. The two countries' figures measure different things and are not meant for size comparison.

Read the table down the columns and the differences show; read it across the rows and the resemblance shows. The difference is the vessel. One side created a legal entity with shareholders; the other laid a statutory duty onto an existing platform. The resemblance is in the second row from the bottom. In both, the side carrying the quality duty is the same side that gathers the data.

The four sections above are what could be confirmed; from here on is what this article wanted to ask. Once the state gathers and supplies training data, two questions follow immediately. One is judging the quality; the other is answering for the defects once they surface.

5.1The side that supplies doubles as the side that appraises

The Chinese answer is already written into the founding documents. The second of the three clauses of the business remit quoted in Section 1 was industry data governance services. The company that supplies the data also sells, as a product, the work of ordering and managing that industry's data. A structure in which the supplier reviews its own supply is there at the level of the company's list of business lines.

The Korean answer is in the statute. The parties to take the measures necessary for quality management of linked and provided training data are the Minister of Science and ICT and the heads of the relevant central administrative agencies. The side that receives and gathers the data carries its quality too. The two designs differ in vessel and leave the same cell blank. Neither has a verifier separate from the supply.

This is not to say Korea has no framework for measuring quality at all. The Ministry of the Interior and Safety runs a public data quality management level diagnosis and evaluation. Its indicator set is eight indicators across three domains, and in 2025 it covered 7,416 databases at 684 institutions.

Split the eight indicators by what each one actually looks at and the weight falls heavily to one side. Five of them look at the quality management system: how the database under evaluation is managed, how standards, structure and linkage have been set up, whether there are procedures that head off errors in advance. Two look at the data values themselves, namely quality diagnosis and error rate, and the remaining one looks at whether anything was done about the diagnosis results. Five of the eight cells go to how management is being carried out.

In the separately run public data operational status evaluation, in the same year, institutions that scored well on management systems scored poorly on the items covering data AI can actually use. That scheme has a different owner and a different scope, though. It is the Ministry of the Interior and Safety's framework for public data, whereas AI Hub, where AI training data is accumulated, is a programme of the Ministry of Science and ICT and the National Information Society Agency. A figure of 300 standard datasets provided under Interior Ministry management sometimes circulates in the same sentence as AI Hub, but those are numbers from different schemes.

5.2Accumulating and selecting are two different processes

The clearest answer on record to that question is not Chinese but Korean. As the Ministry of Science and ICT stated on 13 October, AI training datasets built through government-funded projects numbered 945 as of the end of 2025, or 955 counting the ten under construction this year. The supply-side number has grown year after year. It is worth knowing, though, that the unit itself, the count of datasets, wobbles from record to record. We confirmed earlier a case in which a press release and the distribution portal wrote different counts for the same data.

Then in May 2026 the government ran a separate programme. It reopened all 691 datasets accumulated in AI Hub through 2022, and, judging by room for extension toward generative AI and by actual use, selected 30 for rework after outside expert review. The budget was $2.1M (KRW 3 billion). A distinction should be kept straight here. Thirty is the number picked, not a verdict that the remaining 661 are unusable. No document carries such a verdict, and the possibility that a $2.1M budget constrained how many could be picked has to be weighed alongside it.

What that programme shows is simple. Accumulating and selecting are different processes, and they use different budgets, different procedures and review by different people. Putting already-accumulated data to a new use meant reopening every item, and that work carried its own money. Korea has, with its own records, demonstrated once that a rising count of datasets supplied does not pull quality up with it. This is not a Korean failure story but the physical shape of a problem both countries carry. China's supply is far larger in scale, and reopening a supply at that scale costs more.

The liability cell is emptier still. If data supplied by the state turned out to contain material of unclear provenance, or material whose terms of use do not permit what was done with it, what can a company that trained a model on it ask, and of whom? Neither country's documents answer. On the Chinese side the supplier also holds governance, so the party being asked and the party answering are the same; on the Korean side the duty to take quality management measures rests with administrative agencies, which is not a rule allocating liability for defects in the data provided.

What changes when the supply window narrows to a single counter? That question has an answer accumulated not in policy documents but in the academic literature. The core of it is that when a supply chain narrows, risk does not disappear; it becomes correlated. Gaps and biases that were scattered back when everyone scraped their own now become a shared failure across models the moment they all draw on the same material.

Condensed into six rows, that accumulation holds three strands. Two studies directly measured whether outcomes converge among parties sharing the same component; two audits counted how traceable and how available the provenance of today's data actually is; and two experiments measured what changes when accumulated material is fed back or filtered. The columns are split in two: what each study actually measured, and the range it does not reach.

Source: the papers themselves. The "measured" column carries the figures the papers report; the "limits" column combines the scope the authors state for themselves with how far the work reaches on this article's subject.

Do not walk past the limits cell in the second row. The monoculture field evidence came out of hiring and allocation, and no study yet proves the same thing happens in the pretraining corpora of large language models. When these studies get pulled into a conversation about a national supply counter, jumping to "draw on the same data and the models break the same way" overstates them. What can be said so far is this. The more components are shared, the more outcomes were observed to converge, and sharing training data was the clearest axis of the three.

Fairness also requires putting the fact that runs the other way on the table. The contrast "a state counter concentrates, the private market disperses" is wrong. Of 47 language models released between 2019 and 2023, more than 64% were trained on filtered Common Crawl, and GPT-3 drew more than 80% of its training tokens from it. The whole industry has been leaning on a copy of the web scraped by a single non-profit. Concentration is not a phenomenon the state created by opening a counter; it was already the condition. A state counter changes who owns that concentration.

At the other end of the spectrum sits Europe. The EU Data Act appeared in the Official Journal in December 2023, most of its provisions began to apply on 12 September 2025, and from 12 September 2026 a requirement to ensure access by design was added. Yet as of September 2026, only six of the 27 member states had finished designating their competent authorities. Europe wrote access rules instead of standing up a supplier, and has not yet seated everyone who is to enforce those rules. Found a company, lay down an obligation, or write the rules: all three approaches stall in the same place. Who takes on the work of measuring whether the data is any good is left blank.

Sections 1 through 6 are what could be confirmed from the two countries' official documents and announcements and from the academic literature. This section is the part those documents do not say, and it is about why Pebblous spent a long time looking at this event.

7.1Pebblous does not make data

Put the two designs side by side and one thing they share stands out. The side that gathers and supplies the data was designed with care, while the work of seeing whether the gathered data is any good was left to the same hands. China handed the supplier the industry data governance service as well; Korea laid the quality management duty on the ministries that receive and gather the data. What Pebblous does with AI-Ready Data and DataClinic sits exactly in that unseparated place. We do not make the data; we only look at whether the data someone made is fit to use.

7.2Narrow the supply chain and the gaps get shared too

This event is unusual in that a conversation about data quality, which used to sit at company level, has moved up to the level of a country. When many models train on a corpus drawn from one counter, that corpus's gaps, biases and duplicates become a shared failure across those models. As Section 6 showed, there is already an observation that outcomes converge most clearly when training data is shared. But the setting of that observation was hiring and allocation, and there is still no proof that the same happens in pretraining corpora. So we do not assert it here. We turn it into a measurable question instead: when a national counter ships a dataset, do its provenance, its terms of use and its missing-data rate ship with it?

7.3April 2027: the homework lands on practitioners

Korea's amendment takes effect in April 2027. From then on, institutions carrying out government-funded projects must hand their output data to AI Hub, and must be able to explain what state that data is in. Diagnosing the condition of data before handing it over becomes a new practical task. Korea has already paid for that work once. Reopening all 691 datasets accumulated through 2022 and picking 30 for rework cost $2.1M (KRW 3 billion). What is not recorded while the data is being accumulated gets paid for later by whoever has to reopen all of it.

Companies looking to bring in and use Chinese data carry a different piece of homework. How far the provenance and the terms of use of the datasets the National Data Group ships can be traced becomes a precondition. Recall the audit results from Section 6 and this worry is not vague. Even among widely used public datasets, licence information was missing in more than 72% of cases.

7.4The verifier's seat is still empty

Once the state gathers the data, who looks after its quality? No design yet writes an answer to that question. Not the side that founded a company, not the side that laid down an obligation, not the side that wrote the rules. That it is blank also means the seat is opening up at an institutional level. Pebblous stands not as a competitor to the suppliers but as a verifier separate from the supply. We stay out of the race to gather more data and stand on the side that looks, from outside, at whether the gathered data is fit to use.

The sources differ in grade, so they are grouped. The first group is policy documents and statistical reports; the Chinese figures and the Korean statutory text in the body come from here. The second group is press coverage: the National Data Group's shareholders and leadership were obtained by cross-checking this group rather than from the original notice, and the body says so. The third group is academic literature, and the limits column in Section 6 carries the scope the authors state for themselves.

Policy and statistics (primary)

• 1.SASAC (国务院国资委). Announcement No. 1 of 2026 — approval of the establishment of National Data Group Co., Ltd. ⚠️ The original could not be opened directly, so it is cited via media coverage, and the body says so.
• 2.National Data Administration (国家数据局). National Data Resources Survey Report (2025) 《全国数据资源调查报告（2025年）》 (2026-04-29) — the production total of 52.26ZB, storage total of 2.53ZB, 199.48EB used for training and inference, and 37.84PB of public data opened, all in the Section 3 table. The company-produced share of 33.54ZB (64.18%, up from 58.30%) written after the table, and the rise in authorised-operation data volume from 4.93PB to 7.59PB in Section 2, come from the same report. nda.gov.cn
• 3.National Data Administration (国家数据局). China Data Industry Development Report (2026) 《中国数据产业发展报告(2026)》 (2026-08-28) — data industry size of RMB 6.78tn, 482,000 data companies, 126,000 cumulative high-quality datasets at 1,815PB. nda.gov.cn
• 4.National Data Administration (国家数据局). Results announcement for the "Data Elements ×" three-year action plan (2024–2026) (2026-06-25) — the plan period and "wrap-up year" description in Section 2, and the plan's two targets (at least 300 application cases by the end of 2026, data industry growth above 20% a year on average). news.cn
• 5.National Data Administration (国家数据局) announcement (2026-08-25) — the statement in Section 2 that authorised operation of public data (授权运营) had begun in earnest in 25 of 31 provincial-level administrative divisions.
• 6.Shanghai Data Research Institute (上海数据研究院). Blue Book on the Development of Data Groups 2025 《全国数据集团发展蓝皮书2025》 (as of 2025-09) — the 29 or more provincial groups, 187 prefecture-level groups and 64% of prefecture-level cities in the Section 2 diagram.
• 7.Ministry of Science and ICT press release (2026-10-13) — Cabinet approval of the AI Framework Act and Data Industry Act amendments. Section 4's covered obligation, exemption grounds, the parties carrying the quality-management duty and the April 2027 effective date, plus Section 5's cumulative 945 datasets and ten under construction, come from here.
• 8.Ministry of the Interior and Safety. Public data quality management level diagnosis and evaluation — Section 5's eight indicators across three domains, the 7,416 databases at 684 institutions in 2025, and the 300 standard datasets provided. The breakdown of the eight indicators (quality management system 5 · data value management 2 · action on diagnosis results 1) is from the same notice. mois.go.kr
• 9.Regulation (EU) 2023/2854 (EU Data Act) and Bird & Bird, "The EU Data Act: Where Things Stand Now" (2026-09) — Section 6's application dates and the statement that only six of 27 member states had designated competent authorities. twobirds.com

Press (used with attribution stated)

• 10.Xinhua, Korean-language edition. Launch of the National Data Group (2026-10-22) — the verbatim business remit in the Section 1 quote box, which the body gives as an English rendering of that Korean text, and the description of Xiong'an as proving ground and incubator. This is the only core sentence this article fixes as state-media verbatim. The article contains no figure of 100 central SOEs in total. kr.xinhuanet.com
• 11.The Paper (澎湃新闻). "国家数据集团有限公司组建成立，位列央企名录第28位" — Section 1's No. 28 on the register, the total of 100, the seven shareholders and the leadership backgrounds. Cross-checked against Tencent, Sina Finance, Huxiu, ifeng and Sohu coverage. thepaper.cn
• 12.Money Today (2026-10-21) · The Hankook Ilbo (2026-10-21) — Korean coverage of the launch. That neither article carries the RMB 2 trillion figure was used as grounds for the statement in Section 3. mt.co.kr
• 13.Bloomberg. "China Prepares $295 Billion Plan to Fund Nationwide AI Buildout" (2026-06-09) — Section 3's RMB 2 trillion, the unified computing network by 2028, 80% domestic chips, RMB 5 trillion including power, the six networks, and roughly $725 billion of US big tech spending in 2026. The body states that this is reporting on a draft. Cross-checked with follow-up analysis in Tom's Hardware. bloomberg.com
• 14.Xinhua (新华网). "2025年我国年度数据生产量达52.26ZB" (2026-08-07) — the growth figure in the Section 3 table follows this article's wording verbatim (同比增长27.3％). news.cn
• 15.Korea Data Economy News · ZDNet Korea (2026-05) — Section 5's full review of 691 datasets, the selection of 30 for rework and the KRW 3 billion budget. ⚠️ The framing in the headline, that "only 30 are usable," is not adopted in the body. What the article text says verbatim is that 30 were finally selected after a full review and outside expert review, and no document judges the rest unusable. dataeconomy.co.kr
• 16.Newspim (2026-10-13) · The Busan Ilbo (2025-09-30) — Section 4's explanation of the amendments and the headcount figures at the time of the National Data Agency's elevation (689 to 699, a National Data Management Bureau of 102, and the complaint about four working-level staff). newspim.com
• 17.Newspim · Global Economic (2026-10-21) — Section 2's registered capital of RMB 5 billion for Shanghai Data Group and RMB 10 billion for Fujian Big Data Group.

Academic

• 18.Bommasani, R. et al. "Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?" NeurIPS 2022 — the shared-component hypothesis and the effect of sharing training data, in the first row of the Section 6 table. arXiv:2211.13972
• 19.Bommasani, R. et al. "Algorithmic Monocultures in Hiring" (2026) — the field evidence covering 3 million applicants and 4 million applications in the second row of the Section 6 table. The limit, that the setting is hiring and allocation, is stated alongside it in the body. arXiv:2605.27371
• 20.Longpre, S. et al. "The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI" (2023) — the missing licence information above 72% and error rate above 50% in Sections 6 and 7.3. arXiv:2310.16787
• 21.Longpre, S. et al. "Consent in Crisis: The Rapid Decline of the AI Data Commons" NeurIPS 2024 — Section 6's blocking of more than 28% of C4's core sources and nearly 45% of news domains. arXiv:2407.14933
• 22.Shumailov, I. et al. "AI models collapse when trained on recursively generated data" Nature 631, 755–759 (2024) — model collapse in Section 6. doi.org
• 23.Tirumala, K. et al. "D4: Improving LLM Pretraining via Document De-Duplication and Diversification" NeurIPS 2023 — Section 6's 20% training efficiency gain and up to 2% average improvement across 16 tasks.
• 24.Common Crawl's own About page and a 2024 Mozilla Foundation study — the counter-example in Section 6. The figures that more than 64% of the 47 language models released between 2019 and 2023 were built on filtered Common Crawl, and that GPT-3 drew more than 80% of its training tokens from it, are carried via coverage citing that study. commoncrawl.org

Related Pebblous articles

• 25.Who decides whether one of the amendment's four exemptions applies (Section 4)
• 26.A public data evaluation where good management still left too little data AI can use (Section 5)
• 27.A case where a press release and a distribution portal counted the same training data differently (Section 5)
• 28.The country that counts tokens the way it counts GDP (Section 3)
• 29.Where "opened" and "usable for training" part ways (Section 5)