China's East Data West Compute plan: hubs, capacity and what is reported

China's East Data, West Computing plan moves data processing from crowded eastern cities to inland regions with cheaper land and power. It is built around eight national computing hubs and ten data centre clusters. By the end of June 2026, the industry ministry put national intelligent computing capacity at 2,185 EFLOPS, up 177 percent in a year.

This post explains the plan in plain terms, sets out the figures published by Chinese officials and state media, adds one outside analyst's view, and says plainly where the evidence stops. Each number carries its date and source.

Key numbers at a glance

• 1.88 million PFLOPS, FP16 (National Data Administration head Liu Liehong, Digital China Summit, 28 to 29 April 2026): total intelligent computing capacity as of the end of March 2026.
• More than 80 percent (same speech): the share of that capacity located in the eight national hub nodes.
• 1.37 million PFLOPS, about 72 percent of the total (same event): the capacity connected to the national monitoring and dispatch platform.
• 2,185 EFLOPS (Ministry of Industry and Information Technology, reported by China Daily on 15 September 2026): intelligent computing scale at the end of June 2026, up 177 percent year on year.
• 9,800 EFLOPS by the end of 2030 (MIIT blueprint, same report): the national target, alongside 3.8 trillion yuan of cumulative information infrastructure investment.
• More than 70 (MIIT, same report): computing transmission channels linking hubs and key regions; 17 regional interoperability nodes approved.
• More than 7 GW (China Mobile general manager Chen Yangfan, same report): China Mobile's contracted AI data centre scale. A company figure, not a national one.
• 43.5 billion yuan, about US$6.1 billion (Xinhua, 30 August 2024): direct investment in building the eight hubs by the end of June 2024.
• More than 200 billion yuan (same report): total investment the hubs' construction had driven by then.
• More than 1.95 million (same report): data centre racks, as of mid 2024.
• 20 milliseconds (same report): the network delay between eastern and western hubs, which Xinhua said generally met the requirement.
• 1.04 (same report): the lowest power usage effectiveness reported for newly built data centres.
• 8 hubs and 10 clusters (Xinhua, 2024): the design of the programme, launched in 2022.
• 400 to 500 billion yuan (analyst citing the planning agency, May 2024): the official investment estimate for the whole plan, according to a secondary source.
• About 239 billion yuan (same analyst's own tally of reported projects, May 2024): what the author counted, which is not an official figure.

The problem the plan tries to solve

China's biggest cities and economic regions are in the east. Data centres, the buildings full of servers that run websites, apps and AI models, tend to be built near those users. But land and electricity in eastern cities are expensive and increasingly tight. Inland regions have more land, and some have abundant energy and cooler climates, which can lower cooling costs.

The "East Data, West Computing" idea, in Chinese Dongshu Xisuan, is to send suitable work westward. It was launched in 2022 by the National Development and Reform Commission, according to a Construction Briefing report of 2 September 2024. The design pairs national computing hubs, which are planning areas, with national data centre clusters, which are the concrete groups of facilities inside them. Reports describe eight hubs and ten clusters. A March 2026 China Daily report lists the hubs as Beijing, Tianjin and Hebei; the Yangtze River Delta; the Guangdong, Hong Kong and Macao Greater Bay Area; Chengdu and Chongqing; Inner Mongolia; Guizhou; Gansu; and Ningxia. The first three are the wealthy eastern regions where demand sits. The other five are inland.

What work goes west, and why latency matters

Latency is the delay between sending a request and getting a reply. Real-time services such as payments, video calls or a chat answer need low latency. Batch work such as storage backup, offline analysis and model training does not.

An outside analyst, Andrew Stokols, writing on 28 May 2024, quotes the 2021 planning document as focusing the hubs in Guizhou, Inner Mongolia, Gansu and Ningxia on non-real-time needs: background processing, offline analysis and storage backup. He paraphrases latency goals of about 10 milliseconds one way for the core eastern regions and about 20 milliseconds for remote ones. That paraphrase is not an official quote. Xinhua's 2024 report says delay between eastern and western hubs generally met the 20 millisecond requirement.

This is why the plan places training work and storage in the west more readily than live user-facing services. Training a large AI model is a long batch job, and it can run where power is cheaper, then the finished model can be served from places closer to users. None of the sources cited here breaks down how much AI training specifically runs in the western hubs, so treat that as a design intention, not a measured fact.

What the official numbers say

Mid 2024: spending and scale

Xinhua reported on 30 August 2024 that China had directly invested more than 43.5 billion yuan in building the eight hubs by the end of June 2024, and that the hubs' construction had driven more than 200 billion yuan of total investment. Racks exceeded 1.95 million. A rack is a standing cabinet that holds servers, and it is a common unit for counting data centre capacity, though racks differ in power draw.

The same report said newly built data centres had reached a power usage effectiveness of as low as 1.04. PUE is the ratio of a facility's total energy to the energy used by the computing equipment alone. A PUE of 1.0 would mean no energy lost to cooling and other overhead. Lower is better, and 1.04 is a best case for new sites as reported, not a national average.

Liu Liehong spoke at a big data expo in Guiyang, Guizhou. According to Xinhua he said eastern computing demand was increasingly being met in the west, and that cities would be encouraged to pilot different technical routes over the next two to three years. The report also noted that China aimed to have a preliminary, comprehensive computing infrastructure system by the end of 2025. No official assessment of whether that goal was met appears in the sources cited here.

April 2026: capacity, dispatch and a change of emphasis

At the ninth Digital China Summit in Fuzhou, held on 28 and 29 April 2026, Liu Liehong said China's computing infrastructure was moving from "resource expansion to system-level efficiency gains," according to CGTN's report of 30 April 2026. He described a shift from supplying resources to delivering task-based services, with a goal of computing power that is widely accessible and available on demand.

He listed East Data, West Computing as a priority, along with nationwide monitoring and dispatch of computing resources, coordination between computing and power systems, and edge computing. The summit released a comprehensive report on the national integrated computing power network and a report on coordinated development of computing and networks.

As of the end of March 2026, total intelligent computing capacity was 1.88 million PFLOPS in FP16, with more than 80 percent located in the eight hub nodes. The national monitoring and dispatch platform had connected 1.37 million PFLOPS, about 72 percent of the total.

September 2026: a new number and a 2030 target

At the 2026 China Computational Power Conference in Langfang, Hebei, the Ministry of Industry and Information Technology (MIIT) said intelligent computing scale had reached 2,185 EFLOPS by the end of June 2026, up 177 percent year on year, according to China Daily on 15 September 2026. One EFLOPS is a thousand PFLOPS, so this is 2.185 million PFLOPS. The figures line up: the end of March total was 1.88 million PFLOPS, and a March 2026 China Daily report cited the National Data Bureau putting intelligent computing at 788 EFLOPS as of the end of June. That report gives the month without the year; read as June 2025, it matches the 177 percent growth MIIT gave.

The same report described a new MIIT blueprint with a target of 9,800 EFLOPS by the end of 2030 and 3.8 trillion yuan of cumulative information infrastructure investment over the period. It said more than 70 computing transmission channels linking hubs and key regions had been built, and that 17 regional nodes for national computing interoperability had been approved. These are official statements as reported by state media, and the target is a plan, not a forecast of what will be built.

The nearest thing to a power figure in these reports comes from a company. China Mobile general manager Chen Yangfan said the operator's contracted AI data centre scale exceeds 7 GW, and the report linked those facilities to East Data, West Computing. That is one operator's contracted figure, not a measure of national consumption.

The problems officials and advisers name

Not every official voice describes the plan as a success. In a China Daily report of 12 March 2026, Lian Yuming, a member of the national committee of the Chinese People's Political Consultative Conference, an advisory body, called for the strategy to be recalibrated. He listed several problems:

• A mismatch between supply and demand, with a surplus of low end general computing and a shortage of high end intelligent computing.
• Low utilization in western regions alongside a supply crunch in the industrial east.
• Computing islands, caused by inconsistent technical standards and market rules that block resources from flowing between regions.
• High network latency and transmission costs, which weaken the location advantage of western hubs.
• Weak coordination between computing and electricity, with data centres lagging in clean energy integration.

His recommendations include a three tier structure of national, regional and edge computing, a unified scheduling mechanism and cheaper network fees. These are the views of one adviser, not a government decision, but they match the official emphasis on dispatch and on coordinating computing with power systems.

How to read a computing power number

Chinese official figures are given in PFLOPS, not megawatts. That is a different unit from the one used in most Western discussions of AI data centres, so the two do not line up.

A PFLOPS is a quadrillion floating point operations per second. FP16 says the arithmetic was counted at 16 bit precision. A chip rated at a given FP16 speed is typically rated at a higher figure at 8 bit precision, so any comparison must use the same format. "Intelligent computing" in these reports refers to capacity aimed at AI work, as opposed to general purpose computing.

Three cautions follow.

First, a PFLOPS total is a theoretical peak added up across installed chips. It does not say how busy the machines are, how many sit idle, or how many are old.

Second, it does not say what chips they are. As the explainer on China's domestic AI chips sets out, Chinese chips and imported chips differ in speed and in how well they can be linked.

Third, power is the missing number. None of the official sources cited here gives a national figure in megawatts or gigawatts for China's AI data centres, and PFLOPS cannot be converted to megawatts without knowing the chips. For the power angle in other regions, forecasts are compared in AI data centre electricity demand forecasts compared and an Asia Pacific overview is in APAC AI data centres in 2026.

One outside view of how the money is spent

Stokols's May 2024 analysis is a secondary source, not an official report, but it adds useful context. He names clusters such as Gui'an in Guizhou, Qinyang in Gansu, Zhongwei in Ningxia, Horinger and Ulaanqab in Inner Mongolia, the Chengdu and Chongqing areas, and Zhangjiakou in Hebei.

On money, he cites an official estimate of 400 to 500 billion yuan for the programme, and says his own tally of reported projects reached about 239 billion yuan by May 2024, which he counted as over half the planned scope. He notes that subsidies, such as tax breaks and discounted land, are hard to quantify.

On who builds, his analysis says China Mobile and China Telecom lead reported investment, followed by Huawei, with Alibaba and Tencent comparatively minor in these hubs. He also raises open questions: how much local employment data centres create, how opaque the subsidies are, and whether the Gansu and Ningxia clusters, described as still in their infancy at the time, will attract enough demand. These are one author's views.

How big is one AI computing building?

For a sense of physical scale, Huawei's Connect 2025 keynote described a planned Atlas 950 SuperPoD of up to 8,192 chips, in 160 cabinets over about 1,000 square metres. That is a company claim about a future product, not a hub measurement. It illustrates that very large AI systems can fit in a floor area far smaller than a typical campus, while their power and cooling needs scale with the chips inside.

What is firm and what is uncertain

Firm, because officials or state media report them:

• The structure of eight hubs and ten clusters, launched in 2022.
• The mid 2024 investment, rack and latency figures.
• The end of March 2026 capacity figures from the head of the National Data Administration, and MIIT's end of June 2026 figure of 2,185 EFLOPS.
• MIIT's 2030 target of 9,800 EFLOPS, as a stated plan.
• The stated shift in emphasis toward system-level efficiency, dispatch and power coordination.

Uncertain or missing:

• Utilization rates. No official source cited here reports how much of the capacity is in use, and a national adviser describes utilization in the west as low.
• Power consumption and the energy mix of the hubs. The only gigawatt figure cited here is China Mobile's contracted 7 GW.
• Whether the end of 2025 goal for a preliminary system was met.
• The chip mix inside the 2,185 EFLOPS, and how much of it is domestic.
• Total spending across the whole plan, where an official estimate and an outside tally differ.
• Different sources also name the lead agency differently over time, including the planning commission and the National Data Administration, reflecting restructuring.

What to watch next

• Annual or quarterly capacity updates from the National Data Administration.
• Any publication of utilization or power figures for the hubs.
• Progress in the coordination of computing and power systems that officials named as a priority.
• Whether the dispatch platform's share of national capacity rises above 72 percent.
• Progress toward the 9,800 EFLOPS target for 2030 in MIIT's next updates.
• US export controls, which shape which chips fill these buildings. The timeline of the official texts sets out what each rule says.

If you maintain software, you can check which open source components it contains with the free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• CGTN, 30 April 2026: Three outcomes released at the Digital China Summit
• China Daily, 15 September 2026: China's smart computing scale surges
• China Daily, 12 March 2026: Call to optimize computing power network
• Xinhua via China Services Info, 30 August 2024: China invests over 6.1 billion US dollars in major computing hubs
• Construction Briefing, 2 September 2024: China has invested 6.1 billion in data centre projects
• Andrew Stokols, Sinocities, 28 May 2024: How is China's Eastern Data Western Compute developing?
• Huawei, Connect 2025 keynote, 18 September 2025

Last reviewed: 19 October 2026.