China has ~12% of global AI-relevant compute
… China has managed to maintain about 12% of the world’s AI-relevant compute …
Mid 2026: China Wakes Up · Archived PDF, p. 6
At a glance
- Assessment: Emerging
- Confidence in assessment: 55%
- Outcome: unresolved
- Timing: Original window passed; outcome unresolved
- Evidence: proxy
- Predicted timing: Mid 2026
- Primary source: Mid 2026: China Wakes Up, archived PDF p. 6
What AI 2027 Predicted
Tracker summary: The scenario states that by mid-2026, China has “managed to maintain about 12% of the world’s AI-relevant compute” despite US chip export controls. This is presented in the context of China partially offsetting export restrictions through domestic chip production (Huawei Ascend), stockpiling, and smuggling, while still falling significantly behind the US in aggregate compute capacity. Archived PDF, p. 6.
Scenario source locator: Mid 2026: China Wakes Up: archived PDF, p. 6.
How We Track This
Registered comparison: China’s share of the world’s installed, operational AI-relevant compute as of June 30, 2026, with numerator and denominator normalized using the same documented chip-performance measure and geographic coverage. The tracker interprets “about 12%” (archived PDF, p. 6) as a 10-15% band, inclusive. Report uncertainty and whether independent comparable estimates converge within that band; a range crossing its boundaries is not definitive confirmation. Delivered chips, newly added watts and a selected cluster sample are proxies, not substitutes for the global operational-stock denominator. This band is a tracker interpretation, not a scenario quotation.
We monitor:
- Epoch AI’s GPU cluster and AI supercomputer datasets
- CSET (Georgetown) estimates of Chinese AI compute capacity
- SemiAnalysis reports on global GPU deployment
- Huawei Ascend production volume estimates
- Industry analysts tracking chip smuggling and stockpile estimates
Current Evidence
Epoch’s October 27 AI Chip Imports release models cumulative arrivals from 2022 through June 2026 at 27.4 million H100-equivalents: 19.764 million (72.16%) in the US and 2.637 million (9.63%) in China, including estimated domestic production. The methodology’s October 15 allocation uses a provisional 144-case pilot: China’s conditional 90% interval is 2.26-2.96 million H100e, while its 1.90-4.19 million structural envelope is not a probability interval. Customs allocation and chip-performance normalization do not establish operational stock or actual training FLOPs. Some chips are not yet running or may move elsewhere, and unit shares differ from compute shares. This is provisional context for the registered mid-2026 12% test, not its resolution. Epoch release · Explorer country CSV, retrieved November 7, 2026 (2026 Q2 cumulative H100-equivalents; shares calculated over all listed destinations) · Methodology, sections 4.4-4.5
Epoch AI Data: Epoch AI’s AI Supercomputers dataset (covering roughly 10-20% of total global AI computing capacity) shows the US hosting the majority of GPU cluster performance, followed by China. While this dataset undersamples the full picture, it provides the best available comparative baseline.
Tracked cluster share: Epoch’s June 5, 2025 country comparison reports China at about 15% of represented performance as of May 2025; the previously stated 14.1% was not reproduced from the checked primary page. This is a geographically located, publicly tracked cluster measure, not a census of all Chinese-controlled AI compute. It provides context without resolving the original share. Epoch
Historical compute goal: China’s Ministry of Industry and Information Technology and five other authorities issued the plan targeting more than 300 EFLOPS of total computing power by 2025, according to Xinhua on China’s State Council website, October 9, 2023. This is an aggregate computing-power target, not a measured AI-only deployed stock or its share of a comparable global total.
Private Sector Dominance: Epoch AI notes that industry’s share of AI supercomputer performance in Epoch’s dataset rose from about 40% in 2019 to 80% in 2025. Since US hyperscalers dominate the private sector, this trend may further concentrate compute in the US.
SemiAnalysis deployment-flow estimate (August 2026): Dylan Patel of SemiAnalysis estimated that about 70% of newly deployed AI data-center watts are currently going to the United States and less than 10% to China. This is a specialist analyst estimate based on proprietary industry tracking, not a publicly reproducible census. It measures the flow of new power capacity rather than China’s share of the installed global compute stock, and watts do not adjust for differences in chip efficiency. The estimate therefore suggests downward pressure on China’s global share without establishing that the tracked stock is already below 10%.
Counterevidence & Limitations
- Available data captures only a fraction of global AI compute. Undisclosed military and commercial installations can create unequal coverage across countries.
- The roughly 15% figure is from Epoch’s country comparison, not GeoCoded; it is based on publicly tracked clusters and may overcount or undercount depending on reporting biases.
- China’s actual compute is hard to measure due to: stockpiled pre-ban chips (estimated in the hundreds of thousands of A100/H100-equivalents), smuggled chips, domestic alternatives, and classified military systems.
- The rapid pace of US datacenter construction (Stargate, etc.) may be diluting China’s share even as China’s absolute compute grows.
- “AI-relevant compute” is not precisely defined: different measures (FP32 vs. FP16 vs. INT8; training vs. inference) give different pictures.
What Would Change Our Assessment
- Upgrade to “on-track”: Updated Epoch AI or CSET estimates placing China at 10-15% of global AI compute by mid-2026
- Upgrade to “confirmed”: Multiple independent estimates converging on ~12% for China’s share at the predicted timeframe
- Resolve below target: Comparable installed-stock estimates place China materially below the predicted share. New-deployment flows alone do not establish the stock share.
See Also
- AI 2027 vs Reality: geopolitics category assessment
- Which predictions came true?: full scorecard
Update History
| Date | Update |
|---|---|
| 2026-11-07 | Assessment reviewed. Status: Emerging; confidence: 55%. Epoch’s October 27 chip-imports release estimates roughly 10% of delivered compute in China, while its May 2025 tracked-cluster comparison gives about 15%. These differently scoped proxies bracket the target’s scale but do not form a statistical interval or establish China’s June 30, 2026 operational-stock share. The original window has passed; the outcome remains unresolved. Correction: The June 30, 2026 test concerns installed operational compute and retains the inclusive 10-15% interpretation. Corrected the tracked-cluster attribution to Epoch and the reported share to about 15%, replacing the unverified 14.1%. Linked the imports CSV and distinguished imports, deployment flows and sampled clusters from global operational stock. The 300-EFLOPS policy target is from the October 9, 2023 State Council/Xinhua report; it is not an AI-only measurement. |
| 2026-11-02 | Added Epoch’s customs-based delivered-chip model and both uncertainty ranges. Arrival estimates are not operational-stock shares and do not resolve the mid-2026 12% test; status and confidence unchanged. |
| 2026-09-06 | Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations. |
| 2026-08-31 | SemiAnalysis estimated that about 70% of newly deployed AI data-center watts are going to the US and less than 10% to China. This is an attributed proprietary estimate of deployment flow, not a reproducible measure of China’s installed compute stock. Status and confidence remain unchanged. |
| 2026-03-13 | Best available data suggests China holds ~14% of tracked AI supercomputer performance (2025 data), roughly consistent with the 12% prediction but with high measurement uncertainty. |