Home โ€บ Blog โ€บ China's AI factories
AI Factory Case Studies ยท Part 8 ยท China

Built Under Sanctions: How China Makes AI Factories Without NVIDIA's Best Chips

US export controls cut China off from NVIDIA's most powerful chips. China's response has been to use many more of its own, weaker chips, joined by optics into "supernodes", and to power them with electricity it has in abundance. Alongside that sit a state program that built data center hubs in the west, a ban on foreign chips in state-funded projects, and Chinese labs renting NVIDIA hardware abroad. This case study looks at how the pieces fit, and where they do not.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Huawei hardware figures come from Huawei's own announcements and from coverage of a SemiAnalysis analysis of CloudMatrix 384 (we could not read the original). Export-control dates come from the US Commerce Department's rules as summarised by CSIS and Mintz. "East Data, West Computing" figures come from Reuters and DCD reporting of official data; utilization estimates are contested. Chinese chip bans were reported by Reuters from unnamed sources; the directive has not been published. Malaysia figures come from DCD and reports citing the Wall Street Journal and Financial Times. Figures we found only secondhand are marked โš ๏ธ. See Reading the numbers.

1. Case file

๐Ÿ‡จ๐Ÿ‡ณ
China's AI factory modelHuawei supernodes ยท state data center hubs ยท leased capacity abroad
Main supplierHuawei (Ascend chips)
CloudMatrix 384384 Ascend 910C ยท ~300 PFLOPS BF16
CloudMatrix power~560 kW per system (reported)
Unified memory48 TB (CloudMatrix 384)
Next: Atlas 9508,192 Ascend 950DT ยท Q4 2026
Atlas 950 claim8 EFLOPS FP8 (Huawei)
State programEast Data, West Computing ยท 8 hubs (2022)
PolicyForeign chips barred in state-funded DCs (reported)

Huawei's chip roadmap, announced in September 2025, sets an annual cadence: Ascend 950 (2026), 960 (2027) and 970 (2028). Atlas 950 figures are Huawei's claims, not independent measurements.

2. Timeline

Export controls and China's response, 2022 โ†’ 2028

Solid dots are confirmed by a government or company, hollow dots come from the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

On 7 October 2022 the US first made AI accelerators subject to strict export licensing to China. NVIDIA designed lower-performance chips for the Chinese market; the October 2023 update blocked those too. In April 2025 the US also stopped sales of the H20, NVIDIA's China-specific chip. China pushed in the other direction: in September 2025 its internet regulator reportedly told large tech firms to stop buying NVIDIA chips, and in November 2025 Reuters reported guidance requiring state-funded data centers to use only domestic AI chips, with projects less than 30% complete told to remove foreign chips. For more on the chips themselves, see China's chip stack.

3. The bet: scale out to make up for weaker chips

An Ascend 910C delivers roughly a third of the BF16 performance of NVIDIA's GB200, according to reports of the SemiAnalysis comparison. China's answer is to use about five times as many chips in a system, connect them with large numbers of optical links so they behave as one, and accept that the result uses much more power. CSIS has argued that the original US controls underestimated how far Chinese firms could get by scaling out with many less capable chips.

๐Ÿ’ก A different constraintIn the US case studies, the bottleneck was power: Colossus brought turbines, Hyperion is funding ten gas plants. In China the bottleneck is chips, and power is the resource it can spend. The same trade-off that would be ruinous in Memphis is acceptable where electricity is plentiful and cheap.

4. The supernode

NVL72 vs CloudMatrix 384 vs Atlas 950

Choose a measure. NVL72 and CloudMatrix figures are from reports of the SemiAnalysis comparison; Atlas 950 is Huawei's claim for a system due in Q4 2026 and is shown hatched.

CloudMatrix 384 joins 384 Ascend 910C chips with 48 TB of unified memory and links of 392 GB/s each way per chip, using thousands of optical transceivers. It delivers about 1.7 times the compute of an NVL72 but draws close to four times the power: reported as about 2.3 times less power-efficient per unit of compute. Atlas 950, Huawei's next supernode, scales to 8,192 Ascend 950DT chips in about 160 cabinets, using linear-drive pluggable optics. Huawei demonstrated a 1,024-chip version at WAIC in July 2026.

5. More chips, more power

What it takes to match a cluster

An illustration. Set a target amount of compute and the relative performance and efficiency of the chips available; see how many chips and how much power you need compared with NVIDIA's.

Exaflops of BF16, roughly a large frontier cluster at the top end.
45% โ‰ˆ the reported 2.3ร— gap for CloudMatrix 384.

6. East Data, West Computing

In 2022 China launched a national program, East Data, West Computing, to build data center hubs in the west and north, where land and power are cheap, to serve demand from the wealthy east. Reuters reported eight national hubs at launch and an official estimate of about $63 billion a year of investment. By mid-2024, DCD reported, the hubs had 1.95 million server racks, about 63% of them in use.

The national hubs

A schematic of the eight hubs, not to scale. Tap a hub.

โ† west: cheap power east: users & data โ†’
Tap a hub.

How much of this capacity is actually used is disputed. Reports in 2025 described western centres running at 20โ€“30% utilization, far below the policy goal of more than 60%; a figure attributed to the China Academy of Information and Communications Technology put average utilization of AI computing centres at about 32% โš ๏ธ. One analyst who visited more than 50 Chinese AI data centers reported the opposite, utilization above 80% at many sites โš ๏ธ. A common explanation for the gap is that latency-sensitive work stays near users in the east, leaving western capacity for less urgent jobs.

How busy are they?

Different measures from different sources. They do not measure the same thing.

7. Training abroad

The other answer to export controls is to go where the chips are. Reports citing the Financial Times say Alibaba and ByteDance have trained recent models in Southeast Asian data centers, leasing NVIDIA capacity from non-Chinese operators in places such as Singapore and Malaysia โš ๏ธ. ByteDance announced a $2.1 billion investment in AI and digital infrastructure in Malaysia, and a report citing the Wall Street Journal said it planned about 36,000 B200 chips there โš ๏ธ. Alibaba has opened data centers in Johor, bringing its Malaysian total to five.

Three routes to compute

Domestic chips at homeHuawei Ascend supernodes in Chinese data centers. Required for state-funded projects (reported). Limited by chip supply and efficiency.
Remaining NVIDIA at homeOlder or China-specific NVIDIA chips bought before bans, plus reported smuggling (one estimate: ~140,000 GPUs in 2024 โš ๏ธ). Discouraged by Beijing.
NVIDIA abroadLeased or owned capacity in Malaysia and Singapore. Uses top chips, within export rules as reported, but subject to changes in those rules.

8. Power

China's grid has been adding generating capacity far faster than the US grid, which is why the power penalty of weaker chips is tolerable there. The Ascend approach would face the same constraint as the US sites in places where power is scarce, but in China the limiting factors are chip supply, manufacturing yields and software maturity rather than megawatts. Related reading: the power bottleneck and China's chip stack.

9. Eight bets compared

ColossusAbileneFairwaterRainierMetaGoogleCoreWeaveChina
The betSpeedCapitalNetworkOwn siliconEvery buildingWhole stack earlyRent, with debtMany weaker chips, cheap power
ChipsNVIDIANVIDIANVIDIATrainiumNVIDIATPUNVIDIAHuawei Ascend (+ NVIDIA abroad)
Scarce resourcePowerCapitalPowerPower, waterPowerClean powerCreditChips
Main issuePermitsOne customerUtility gasWellsStranded plantsEmissionsDebtEfficiency, idle hubs
Scale out~5ร— the chips per system
Spend power~2.3ร— less compute per watt
State-ledNational hubs, domestic-chip mandates
Offshore optionNVIDIA capacity in Southeast Asia
VerdictChina shows that export controls change how AI factories are built more than whether they are built. Huawei's supernodes trade efficiency for scale, which works where power is cheap; the state program built capacity ahead of demand, some of it idle; and Chinese labs rent the chips they cannot import. The open questions are how fast Ascend chips improve and how much manufacturing capacity China has to make them. Next: sovereign AI factories, from Stargate UAE to India and Europe.

10. Reading the numbers

From Huawei
384 chips and 48 TB per CloudMatrix 384; Atlas 950 with 8,192 Ascend 950DT, 8 EFLOPS FP8, ~160 cabinets, Q4 2026; chip roadmap 950/960/970 for 2026โ€“28; 1,024-chip demo at WAIC 2026.
Analysis (via coverage)
~300 vs ~180 PFLOPS BF16; ~560 kW vs ~145 kW; 2.3ร— less efficient per FLOP; ~3ร— lower per-chip BF16 for Ascend 910C. From SemiAnalysis as reported; we did not read the original.
Governments
US controls 7 Oct 2022 and 17 Oct 2023 (Commerce, via CSIS and Mintz); H20 halted Apr 2025. Chinese guidance on NVIDIA purchases (Sep 2025) and state data centers (Nov 2025) reported by Reuters from unnamed sources.
Contested โš ๏ธ
Utilization (63% of racks, 20โ€“30%, ~32%, >80% at visited sites); 36,000 B200 in Malaysia; ~140,000 smuggled GPUs in 2024.
Illustrations
The matching model and the hub map are simplified.