China's Total Computing Capacity Reaches 3.1 Million PFLOPS, Second Worldwide; Embodied-AI Data Standards Working Group Established (November 10, 2026)

Event overview ​

On November 10, 2026, the National Data Standardization Technical Committee (NDSTC) held the second "standards week" of 2026 in Xiamen, Fujian. CCTV News, citing the National Data Administration, reported two items disclosed at the event: a national figure on computing capacity, and the establishment of a new batch of data standards drafting groups.

On computing capacity, the official formulation was that "China's total computing capacity has reached 3.1 million PFLOPS, ranking second worldwide." That sentence contains two elements — a total figure and an international ranking — and nothing else: no reference date, no precision basis, no breakdown by capacity type.

On standards, the NDSTC formally established an embodied-AI data standards drafting working group, plus four industry-application standards drafting groups covering meteorology, health, electric power and commerce. The official framing is to "accelerate data-related standards development during the 15th Five-Year Plan period and enable digital transformation across industries."

Read together, the two items point to the same logic: computing capacity has reached the world's front rank, and the next bottleneck shifts from "how much compute exists" to "whether data can be supplied and used in standardized form." Carving embodied AI out as its own data standards group is the most concrete signal of that shift.

Key facts ​

What 3.1 million PFLOPS measures: three figures, three different rulers ​

The "3.1 million PFLOPS" figure is easy to misread in isolation, because it does not use the same ruler as several previously published capacity numbers. Lining up the timeline makes the difference clear.

The first reference point is intelligent computing capacity at the end of 2025. According to the National Data Administration's Digital China Development Report (2025), released in June 2026, China's intelligent computing capacity reached 1.59 million PFLOPS (FP16) as of end-2025, ranking second globally.

The second is total intelligent computing capacity reported in July 2026. At the press conference for the 2026 China International Big Data Industry Expo, the National Data Administration said nationwide intelligent computing capacity stood at 2.4 million PFLOPS, that intelligent computing built out across the eight national computing hubs accounted for more than 80% of the national total, and that 1.4 million PFLOPS had been connected to the national integrated computing network trial platform.

The third is the current figure: 3.1 million PFLOPS of "total computing capacity."

The decisive difference lies in the qualifiers. The first two refer to "intelligent computing capacity," and the 2025 figure explicitly specifies FP16 precision; the current figure refers to "total computing capacity," which typically spans both general-purpose and intelligent compute and is therefore a broader measure. In this release, the official statement does not disclose the precision basis or the composition breakdown of the 3.1 million figure.

One point therefore must be stated plainly: 3.1 million cannot be subtracted from 2.4 million to derive "700,000 PFLOPS added in four months," nor compared against 1.59 million to compute a growth rate. Different denominators, different measured objects and possibly different precision bases make any cross-basis growth extrapolation unfounded. The only confirmable fact is this: on the official total-capacity basis, China's compute totals 3.1 million PFLOPS and ranks second worldwide.

A unit conversion helps convey scale: 1 EFLOPS equals 1,000 PFLOPS, so 3.1 million PFLOPS equals 3,100 EFLOPS, or roughly 3.1 ZFLOPS. This conversion is pure arithmetic and involves no assumption about measurement basis.

Why embodied AI needs its own data standards group ​

Among the newly established groups, embodied AI is the only one named for a technology direction; the other four are named for industries (meteorology, health, electric power, commerce). That naming difference is itself telling: what embodied AI lacks is not a data specification for some industry, but a common way to describe data at all.

By common industry understanding, embodied-AI training data differs fundamentally from that of conventional large models. Language models consume text, image models consume pixels, but embodied AI needs "action + perception + outcome" triples: where the arm is, in what pose, with how much force, and what changed in the object. Such data must be collected on real machines or generated in high-fidelity simulation — costly, limited in scale, and described differently by every vendor in terms of capture hardware, coordinate conventions, action-space definitions and annotation formats.

The direct consequence of missing standards is that datasets cannot be reused. The same set of grasping data has to be re-annotated for another company's model; the same simulated environment cannot be ported across platforms. This contrasts sharply with the text-dataset paradigm, where data can be freely concatenated and mixed for training.

Establishing a separate drafting group for embodied-AI data standards means regulators are treating "data that can be exchanged and reused" as a precondition for industrialization, rather than waiting for industry to converge on de facto standards. That aligns with the broader direction of data-factor market reform during the 15th Five-Year Plan period.

From working group to national standard: how far the path runs ​

Three layers need to be distinguished, to avoid equating "a working group was established" with "a standard has been published."

The first layer is the working group. What the NDSTC established is a drafting working group, whose task is to organize the drafting of a standard proposal — the earliest stage of standards development.

The second layer is project approval and review. A data-domain standard typically passes through requirements solicitation, project initiation, drafting, public comment and technical review before reaching publication. At its July 2026 press conference, the National Data Administration offered a set of figures: 130 data-domain standards requirements published, 25 national standards and 21 national guiding technical documents advanced to project initiation, and one international standard on privacy protection released. That illustrates how long the chain from requirement to publication is.

The third layer is adoption and implementation. Even after publication, companies must be willing to rework collection pipelines, and platforms must support the standard formats.

This release does not disclose the new groups' member organizations, the titles of standards they intend to draft, project numbers or timelines. What can be confirmed now is only that the organizations exist and the direction is set; the standard texts have not entered publicly visible stages.

Impact and limits ​

First, the computing narrative is shifting from "scale" to "quality of supply." 3.1 million PFLOPS is a scale conclusion, while the standards groups established the same day point at standardization on the supply side. Anyone judging China's compute progress on the headline number alone will miss this thread.

Second, the industrialization threshold for embodied AI is being redefined. Discussions of humanoid robot deployment have focused on hardware cost and motion control; with data standards entering the picture, "can training data be produced and circulated at scale" becomes an equally important variable. That will directly shape the form of adjacent industries such as data collection services, simulation platforms and annotation tooling.

Third, the compute-power coordination pressure remains. At its July press conference, the National Data Administration explicitly noted that "the end of compute is electricity," that it is optimizing power supply guarantees for data centers and advancing "compute-electricity coordination" through direct green-power connections. As capacity keeps rising, power and cooling become harder constraints — a point that also holds in this month's reports on US AI infrastructure hitting a power bottleneck.

Limits to be explicit about: the reference date for the 3.1 million PFLOPS figure, its precision basis (FP16/FP32/FP64), and the split between general-purpose and intelligent compute are all undisclosed; so are the specific standards lists, member organizations and release timelines of the four industry groups and the new embodied-AI group. This site makes no extrapolation on capacity growth rates, international gap multiples, or standards completion dates.

Official sources ​

• CCTV News: Computing capacity ranks second worldwide as China advances data standards for embodied AI and other fields (November 10, 2026)
• China National Radio, republishing CCTV News: Computing capacity ranks second worldwide (November 10, 2026)
• National Data Administration: Transcript of the 2026 China International Big Data Industry Expo press conference (July 28, 2026, including the 2.4 million PFLOPS intelligent-computing figure)
• Xinhua: Full text of the Digital China Development Report (2025) released (June 9, 2026, including the 1.59 million PFLOPS FP16 figure for end-2025)