Chinese domestic AI chips 3 years behind US-Taiwan

What AI 2027 Predicted

By mid-2026, domestically produced Chinese AI chips remain roughly three years behind the US-Taiwan frontier. Different dimensions of hardware progress need separate comparisons. AI 2027

How We Track This

Compare delivered chip performance at matched numerical precision, memory, interconnect and production scale. Announced roadmaps support forecasts; they are not shipping inventories. A per-chip lag is not automatically a lag of the same size in complete systems.

Current Evidence

Epoch’s September 4 original analysis estimates Huawei’s per-chip performance lag at approximately three to four years. Its forward comparison places the scheduled Q4-2026 Ascend 950DT at 1 PFLOP FP8 against the H100’s 1.979 PFLOPS, with similar memory bandwidth. The H100 dates to 2022. The 950DT comparison is prospective, not a claim that it already shipped. Epoch hardware analysis

Huawei’s own roadmap confirms the distinction between planned products and its system-integration strategy. These sources support a substantial hardware disadvantage of approximately the scenario’s size. Huawei roadmap

Counterevidence & Limitations

Epoch also identifies stronger Chinese progress in networking and system scale, including CloudMatrix 384. Combining more chips can partly compensate for weaker individual accelerators, while changing power, cost and complexity. Its production comparisons are estimates rather than audited shipment totals. Some Ascend production uses imported logic dies or memory, so a Chinese-vendor performance comparison does not by itself establish wholly domestic manufacture.

The current analysis corroborates the approximate midyear claim; it is not a direct measurement of an exact three-year lag on June 30. On-track describes that qualified support, with the dated outcome still unresolved.

What Would Change Our Assessment

• Strengthen: Matched, independent tests of delivered hardware consistently support the approximate gap.
• Weaken: Delivered Chinese systems close the relevant cost/performance and production gaps substantially faster than the scenario.
• Clarify: New evidence separates per-chip progress from system performance and manufacturing volume.

Update History

Date	Update
2026-09-07	Added Epoch’s September hardware comparison and Huawei’s primary roadmap. Evidence supports a substantial approximate per-chip lag; scheduled hardware, system integration and production estimates remain distinct.
2026-09-06	Assessment revised from on-track (0.75) to emerging (0.65). Hardware performance, efficiency and production need separate dated comparisons; a uniform three-year lag is not established.
2026-06-01	Added China’s certification of nine domestic AI processors for government procurement as a domestic-chip ecosystem datapoint. This supports continued adaptation, but does not materially change the three-year-gap assessment without independent performance benchmarks.
2026-04-13	CFR analysis (Dec 2025) concludes Huawei is “falling further behind” — its own public roadmap shows the 2026 chip will be less powerful than its current best. Huawei scaling production to 1.6M dies in 2026 (Bloomberg, Oct 2025) but performance gap persists. DeepSeek V4 on Huawei Ascend 950PR is a milestone but required “substantial re-engineering” and contributed to the model’s delays. Trump admin approved H200 exports to China (Dec 2025), suggesting assessment that Huawei isn’t yet competitive enough to justify strict controls. Overall: China CAN run frontier models on domestic chips but with performance trade-offs. 3-year gap assessment remains broadly consistent. Confidence +0.05. Sources: CFR, HKFP/AFP, findskill.ai
2026-03	Huawei’s Ascend 910C significantly lags NVIDIA H200 in performance. Ren Zhengfei publicly acknowledges being ‘one generation behind.’ A ~3-year gap is a reasonable characterization of the current state.
2025-04	US imposes indefinite H20 export license requirement (April 9). H20 was the last Nvidia chip available to Chinese buyers without license. METR June data suggests gap is real — Chinese models lag ~6 months on agentic tasks, partially attributable to compute constraints.