Huawei Unlikely to Catch Nvidia by 2030, Epoch AI Finds

Huawei is unlikely to catch up to Nvidia in AI chip dominance by 2030, according to a detailed roadmap analysis by Epoch AI, even as Chinese AI companies place large orders for its Ascend accelerators and Beijing pushes for semiconductor self-reliance. The Chinese technology giant’s most advanced accelerator, the Ascend 950, delivers roughly half the performance of Nvidia’s H100, a chip that first shipped in 2022. Huawei’s total compute output in 2026 is projected to remain below 4% of Nvidia’s.

That gap matters far beyond corporate rivalry. AI compute capacity now shapes which countries and companies can train the most advanced models, run large-scale inference, and build independent AI infrastructure. For global executives, suppliers, and policymakers, the Huawei-Nvidia comparison is a real-time measure of whether US export controls are slowing China’s AI ambitions — and how long Nvidia’s technological and ecosystem lead can hold.

The performance and production gap

Epoch AI’s analysis, titled “Huawei’s Roadmap to 2031,” finds that Huawei trails Nvidia significantly in both per-chip performance and total production volume. The Ascend 950, Huawei’s flagship AI chip, offers approximately half the performance of Nvidia’s H100. That is a striking lag because the H100 was introduced in 2022, meaning Huawei’s current leading product is several generations behind Nvidia’s already shipping hardware.

The production gap is even larger. According to the analysis, Huawei’s total compute output in 2026 will be less than 4% of Nvidia’s. This is not simply a matter of design capability; it reflects constraints across manufacturing capacity, advanced packaging, and access to critical components. Even if Huawei executes its domestic roadmap perfectly, the sheer scale of Nvidia’s output — supported by TSMC’s advanced process nodes and a global supply chain — remains difficult to match.

High-bandwidth memory is the near-term bottleneck

A central constraint is high-bandwidth memory, or HBM, which is essential for AI accelerators to move data quickly enough for large-model training and inference. Epoch AI estimates that even with domestic HBM production ramping up, Huawei’s HBM supply could limit its compute output to around 1% of Nvidia’s through 2028.

Smuggling additional HBM into China would help, but only to a point. The analysis calculates that even a tenfold increase in HBM availability through unofficial channels would raise Huawei’s output to about 11% of Nvidia’s by 2028. By then, the binding constraint shifts from memory supply to per-chip performance. Nvidia’s advantage in compute per gigabyte of memory grows from nearly 3× to 9× over that period, meaning Huawei would need not just more memory but fundamentally more efficient chip designs.

DeepSeek’s order shows demand, not parity

Despite these limitations, demand for Huawei chips is rising. DeepSeek, the Chinese AI lab that gained global attention for efficient model training, plans to deploy at least 160,000 Ascend 950DT chips in a new data center in Inner Mongolia, according to The Economic Times. That would be one of the largest known Huawei clusters to date and a significant commercial vote of confidence in domestic silicon.

However, the order does not signal technological parity. DeepSeek still relies on Nvidia accelerators for model training, indicating that Huawei’s chips are not yet competitive for the most demanding AI workloads. The Ascend 950DT deployment is likely aimed at inference or less compute-intensive tasks, where Huawei’s hardware can be useful even if it cannot match Nvidia for frontier training runs. For international observers, the DeepSeek order illustrates that Chinese firms are building a parallel AI infrastructure, but it remains dependent on older-generation Nvidia performance levels.

Nvidia’s roadmap widens the distance

Nvidia is not standing still. Its Rubin architecture claims 5.4× the performance per megawatt over the current Blackwell generation, and the Feynman generation is expected in 2028 on a 1.6 nm-class process. These advances push the frontier further out, raising the bar for any competitor trying to close the gap.

Huawei’s long-term hope is a technique called “LogicFolding,” which stacks logic layers vertically to increase density and performance. But the first Huawei chip using LogicFolding is not expected until 2030 — two years after Nvidia’s Feynman is projected to launch. Even if LogicFolding works as intended, Huawei would still be competing against Nvidia’s 2028 architecture with a chip that arrives later and must overcome the same memory and manufacturing constraints that have held it back for years.

For international executives, the picture is clear: Nvidia’s technological lead remains resilient, and China’s semiconductor self-reliance under US export controls is progressing but still years behind. Huawei is building a credible domestic alternative, and its chips will power a growing share of Chinese AI inference. But by 2030, the analysis suggests, Huawei will not have caught up to Nvidia in the high-end AI chip market that matters most for frontier model training and global AI leadership.