Epoch Says Huawei Is Unlikely to Catch Nvidia by 2030

Huawei is the standard answer to the question of who can build AI chips outside the reach of US export controls. Epoch AI took the company's public roadmap seriously and still concludes it is unlikely to catch Nvidia by 2030. The starting gap is large: Epoch estimates Huawei will produce under 4 percent as much AI compute as Nvidia this year, and its Ascend 950 delivers about half the performance of the H100, a chip Nvidia shipped back in 2022. That is roughly a four-year lag in raw capability.

Three constraints keep the gap open. The first is the process node. Export controls hold Huawei's fabrication partner SMIC to 5 to 7 nanometer-class processes while Nvidia buys TSMC's more advanced nodes, which Epoch reads as about a 2x transistor-density edge that persists through 2030. The second is memory. High-bandwidth memory sets the production ceiling, and if Huawei relies only on domestically made HBM its output could stay near 1 percent of Nvidia's through 2028. The third is timing on the long bet: Huawei's move to vertically stacked logic, which it calls LogicFolding, is not expected in Ascend parts until 2030, while Nvidia's Feynman generation brings comparable stacking in 2028 on better silicon. Epoch's finding is not that Huawei stalls, but that Nvidia moves faster on every axis at once, and by 2028 the per-chip performance gap widens to about 9x. You can read the full report for the assumptions behind each number.

Why it matters

If you are betting on Chinese compute as a hedge against export controls, or pricing the risk of a bifurcated hardware market, this puts numbers on how long the gap lasts rather than whether it closes eventually. For the rest of the decade, access to leading Nvidia and TSMC output stays the deciding factor in who can train the largest models.