Huawei remains light-years behind Nvidia, with the gap widening through 2030

The numerical gap is already here. Huawei's most powerful chip, the Ascend 950, delivers roughly half the performance of Nvidia's H100, a chip that began shipping in 2022. The production-volume picture is worse: this year Huawei will manufacture less than 4% of Nvidia's total compute capacity. If it relies solely on Chinese-made HBM memory, that share could stall around 1% by 2028.

The immediate bottleneck is memory. The constraint isn't logic but HBM supply. Even after blending foreign inventory, lower-bandwidth alternatives, and smuggled volumes, the usable-memory supply gap stands at 73-to-1 in Nvidia's favor in 2026. Smuggling ten times China's entire domestic HBM supply in 2028 would still bring Huawei's compute output to only about 11% of its American rival's.

The bottleneck shifts to performance around 2028. The HBM shortfall is expected to narrow to an 11-to-1 gap, at which point the core problem moves to per-chip performance. Nvidia's compute-per-gigabyte advantage grows from nearly 3x to 9x, while its new Rubin architecture shows a 5.4x performance-per-watt improvement over Blackwell, and the Feynman generation is slated to push that further in 2028.

The long-shot bet: LogicFolding. Huawei's answer to the performance lag is LogicFolding, a vertical stacking of logic to increase transistor density despite being stuck on 5-7 nanometer processes. The first chip to integrate the technology isn't expected before 2030, and the roadmap the company presented at IEEE ISCAS 2026 shows a three-to-four-year lag persisting at least through the end of the decade.

The bottom line: at this rate, the nominal per-chip performance gap reaches 17.5x as early as next year. Export controls haven't stopped Huawei entirely, but they have turned the chase into a slow, expensive, and far-from-sufficient pursuit.