Epoch AI, a leading AI research institution, released a special report on September 4 comparing Huawei and NVIDIA chips. The core conclusion is that due to export control restrictions, it is nearly impossible for Huawei to catch up to NVIDIA by 2030; If Huawei cannot access overseas memory, its AI computing power in 2028 will only be about 1% of NVIDIA’s.
The report titled “Will Huawei catch up to Nvidia by 2030?” highlighted the current performance and production volume gaps between Huawei and NVIDIA chips. It estimated that Huawei’s AI computing power in 2026 will be less than 4% of NVIDIA’s, and without access to overseas high-bandwidth memory (HBM), this data will drop to around 1% by 2028.
The report first analyzed the extent of the gap between Huawei and NVIDIA in 2026 in terms of single-chip performance and chip production volume, with HBM supply being a critical factor limiting Huawei’s output.
Huawei’s current flagship chip, the Ascend 950DT, is one of the most powerful AI chips. The analysis in the report indicated that when the Ascend 950DT performs AI calculations using FP8 (8-bit floating point), its theoretical calculation throughput is approximately half of NVIDIA’s H100, a data center-level chip released by NVIDIA in 2022.
Looking solely at the “theoretical peak computing power” in chip specifications, NVIDIA’s B300 is approximately 6.75 times that of Huawei’s Ascend 950DT. B300 belongs to NVIDIA’s new generation Blackwell Ultra series of data center AI acceleration chips primarily used for training and inference of large AI models.
The report concluded that Huawei’s chip performance currently lags behind NVIDIA’s by approximately 4 years, and it is projected to remain 3 to 4 years behind by 2030. Even if Huawei were to acquire overseas HBM through illicit means, assuming a supply 10 times that of domestically produced HBM, by 2028 Huawei’s computing power will only be around 11% of NVIDIA’s.
One of Huawei’s competitive strategies is to compensate for the deficiency in single-chip performance by connecting more AI chips (logic stacking) and expanding the system scale. However, the report argued that this approach cannot fully bridge the gap in single-chip performance between Huawei and NVIDIA.
A significant bottleneck for Huawei currently is the lack of sufficient HBM for chip production. China primarily relies on vendors like CXMT and XMC for HBM production, but the technological level and capacity of Chinese HBM lag behind that used by NVIDIA sourced from SK Hynix, Samsung, Micron, and other manufacturers.
By 2028, Huawei’s major bottleneck is expected to transition gradually from “insufficient HBM” to “inadequate single-chip performance.”
In terms of chip manufacturing processes, Huawei mainly depends on SMIC. The report indicated that SMIC’s current mass-produced advanced processes are around N+3, 7-nanometer level, while TSMC is already advancing towards 3-nanometer, A16 (around 1.6-nanometer).
It is estimated that TSMC’s advanced process transistor density is about double that of SMIC, resulting in higher computing power, lower power consumption, and the ability to develop smaller chips with the same performance.
As a result, NVIDIA holds a significant technological advantage in chip design and manufacturing, combining advanced TSMC processes with larger chip areas, more chips per wafer, advanced 3D logic stacking, high-performance HBM, and advanced packaging technologies to continuously improve the computing performance of individual AI chips.
Despite utilizing large chips and multi-chip packaging to enhance performance, Huawei is still challenged to fully bridge the gap with NVIDIA in single-chip performance due to limitations in advanced processes, HBM supply, and some advanced manufacturing equipment, and this gap is expected to widen over time.
Huawei’s hopes lie in LogicFolding 3D stacking technology to boost Ascend AI chip performance. However, according to Epoch AI’s analysis, LogicFolding is not anticipated to be applied to the Ascend chips until 2030. In contrast, NVIDIA’s Feynman, projected for release in 2028, may adopt 3D logic stacking on TSMC A16 processes. The report questions Huawei’s ability to catch up with NVIDIA through 3D stacking to compensate for insufficient advanced processes.
Established in April 2022, Epoch AI is an independent, non-profit research institution focusing on AI development, tracking AI model capabilities, computing power, chips, data centers, and trends in the AI industry.
Despite being founded only four years ago, Epoch AI has garnered a good reputation in the industry and is highly credible and influential in quantitative research areas such as AI computing power, model scale, hardware supply, and AI development trends. Epoch AI’s studies have been cited or utilized by institutions such as Stanford HAI, the UK government, Google DeepMind, OpenAI, and others.