Epoch AI: Huawei Will Field 25x Less Compute Than Nvidia in 2026, Gap Unlikely to Close

Epoch AI · rss · 2026-09-25

Epoch AI compares Huawei and Nvidia's AI chip position, concluding Huawei will stay roughly 3–4 years behind in both chip performance and total compute output through 2030, with the gap almost certainly unclosable.

2026 baseline (in H100-equivalents):

Why the roadmap can't close it: fab partner SMIC, cut off from advanced lithography by export controls, can't match TSMC density (domestic litho likely post-2030). Huawei must rely on bigger chips (cost/defect trade-offs), low-precision architecture tweaks, and LogicFolding vertical stacking that only lands in 2030 — while Nvidia's Feynman generation (2028) stacks superior dies two years earlier, compounding a 2× density edge into 4×. By 2027 the flagship gap widens to 9×; the Ascend 970 (late 2028) only approaches the 2025 B300.

Systems and software: Huawei's 8,192-chip SuperPoDs compensate for weaker chips via larger domains, but key performance metrics are unpublished. Its CUDA counterpart CANN remains years behind — DeepSeek's R2 training attempt on Ascend failed even with Huawei engineers on site, reverting to Nvidia. However, if Chinese labs are forced onto domestic chips at scale, the feedback loop could close the software gap quickly.

Original post →

More from Infra

Infra channel →