LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésSujet · 华为英伟达算力差距Structurel3 min de lectureRédaction DisconfirmAI

Epoch AI Analysis: Why Huawei Is Unlikely to Catch Up with Nvidia by 2030

Epoch AI researchers Venkat Somala and Lynette Bye quantify the compute gap between Huawei and Nvidia at 25× in 2026 and argue that export controls constrain both per-chip performance and production volume, leaving Huawei approximately four years behind through 2030.

2026-09-25 00:04 GMT+8

In the core infrastructure of US-China AI competition—compute capacity—China's leading AI chip designer, Huawei, faces significant headwinds. Epoch AI researchers Venkat Somala and Lynette Bye present a central thesis in their September 24, 2026 report: although Huawei plans to substantially improve its AI compute solutions by 2030, US export controls constrain its most important scaling levers, making it unlikely to catch up with Nvidia.

The report quantifies the current gap using H100-equivalents (H100e) as a baseline. In 2026, Huawei's flagship Ascend 950 chip delivers roughly one-seventh the compute throughput of Nvidia's B300 and half that of Nvidia's 2022 H100, suggesting a four-year lag in per-chip performance. On volume, Huawei is estimated to produce 1.5 million chips in 2026 compared to Nvidia's roughly 6 million. Multiplying performance by volume, Huawei will produce approximately 25 times less total compute than Nvidia in 2026.

Huawei aims to close this gap through two paths. On chip performance, because fabrication partner SMIC cannot pack transistors as densely as TSMC—and advanced lithography imports are restricted—Huawei is pursuing bigger chips, architecture optimizations for lower precision, and vertical stacking called LogicFolding. However, vertical stacking won't reach the Ascend line until 2030, while Nvidia is expected to deploy superior stacking in its Feynman generation by 2028, potentially compounding its transistor density advantage from 2× to 4×.

At the system level, Huawei plans to connect 8,192 chips into SuperPoDs to compensate for weaker individual units. Yet maintaining high-bandwidth communication across thousands of chips presents steep engineering trade-offs. Furthermore, Huawei's software stack CANN remains far less mature than Nvidia's CUDA. DeepSeek attempted to train its R2 model on Ascend chips, Huawei even sent its own engineers in, and still couldn't complete a training run, prompting DeepSeek to revert to Nvidia hardware.

Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Production constraints are equally rigid. Huawei relies on domestic suppliers CXMT and XMC for High Bandwidth Memory, whose ramp-up is slow. The report estimates that even by 2028, Huawei's annual output using domestic HBM would reach around 1.5 million H100e, roughly 1.5% of Nvidia's production. With the stockpile of over 10 million foreign HBM stacks built before December 2024 export controls expected to run out in 2026, the volume ceiling becomes more binding. Even assuming Huawei smuggled ten times its domestic HBM in 2028—an implausibly large flow costing over $40 billion—its output would still reach only about 16% of Nvidia's production that year.

The report concludes that Huawei will likely remain around four years behind in both chip performance and compute production. Export controls hit both volume and performance simultaneously, making the path to parity significantly harder, slower, and more expensive. The authors acknowledge Huawei is one of the world's most capable engineering organizations and may eventually catch up, but available information strongly suggests it will not happen by 2030.

Sources

  • epochai.substack.com↗
Sujet

华为英伟达算力差距

【背景】美国对华半导体出口管制持续收紧,中国本土AI芯片龙头华为与全球霸主英伟达之间的算力差距成为衡量中美科技竞争的关键指标。尽管华为在软件生态上取得突破,但硬件性能与产能的双重瓶颈使其难以在2030年前追平英伟达。

6 étapes2 piècesMis à jour 2026-10-25 23:41

Comment on en est arrivé là

  1. 2025-12-15美国外交关系委员会(CFR)发布报告称,华为AI芯片性能落后英伟达5倍且差距扩大,建议维持出口管制。cfr.org ↗
  2. 2026-09-04Epoch AI研究员Venkat Somala和Lynette Bye发文,量化华为2026年总算力仅为英伟达1/25,预计2030年仍落后4年。x.com ↗
  3. 2026-09-24Epoch AI发布详细报告《How far behind Nvidia is Huawei?》,论证出口管制限制了华为芯片性能和产量两个扩展杠杆。epochai.substack.com ↗

Ce qui reste non résolu

  • 华为CANN 9.2.0软件套件发布后,除TileLang外的其他5个DeepSeek工具能否在昇腾芯片上稳定运行?
  • 在非DeepSeek特定超节点架构(如标准x86主机)上,华为昇腾芯片的兼容性问题何时能解决?
  • DeepSeek是否在2027-2028年正式将主要训练和推理负载迁移至华为技术栈?

Également dans ce fil

  • SemiAnalysis Test: Huawei Ascend Hardware Lags A100 by ~5 Years, DeepSeek Software Near Frontier2026-10-23
Lire le sujet complet →
Continuer la recherche →