[News] U.S. Council: NVIDIA–Huawei AI Chip Gap Could Reach 17× by 2H27, Yet H200 Exports Risk Aiding China

While Huawei’s AI chips have made rapid strides, a new U.S. council report finds that NVIDIA continues to lead in overall AI hardware performance, with that lead expected to continue. According to a report by the U.S. Council on Foreign Relations, an analysis combining publicly available AI chip performance data from both companies with estimates of production capacity suggests that Huawei is not gaining ground. Instead, it is falling further behind, hampered by export controls it has yet to overcome.

The U.S. Council on Foreign Relations says the performance gap between leading U.S. and Chinese AI chips is already wide and is set to grow sharply over the next two years. Measured by TPP, the top U.S. AI chips are currently around five times more powerful than their Chinese counterparts, and by the second half of 2027, NVIDIA’s best AI chips are expected to be roughly seventeen times more powerful than Huawei’s.

The report adds that Huawei is unlikely to deliver a chip exceeding NVIDIA’s H200 for at least the next two years. It notes that Huawei does not plan to introduce a processor with higher performance or memory bandwidth than the H200 until the Ascend 960, slated for Q4 2027, with broad availability likely in 2028.

Production Constraints and Scaling Limits

Beyond the performance gap, the report says Huawei will also lag NVIDIA in production capacity. Even under highly optimistic assumptions—800,000 AI chips in 2025, two million in 2026, and four million in 2027—Huawei’s output would still fall far short, accounting for about 5% of NVIDIA’s total AI computing power in 2025, dropping to 4% in 2026 and 2% in 2027.

The gap is effectively impossible to close, according to the report: even a hundredfold increase in Huawei’s AI chip output by 2027 would still fall short of half of NVIDIA’s production. At the same time, rapidly advancing models are driving exponential growth in China’s AI compute demand, meaning the chip shortage is set to intensify rather than abate.

In addition, the report also notes that Huawei is unlikely to produce large volumes of the multi-rack systems it markets as competitors to NVIDIA’s rack-scale platforms. Huawei has promoted products such as the CloudMatrix 384, which uses 384 Ascend 910C chips, and the forthcoming Atlas 950 SuperPod with 8,192 Ascend 950 chips, as alternatives to NVIDIA’s rack-scale systems built around 72 leading NVIDIA chips. However, the report points out that even under the most aggressive chip production assumptions, Huawei will be unable to manufacture these multi-rack systems in meaningful numbers.

H200 Exports: Risk of Boosting China’s AI Compute Power

While the report states that China is unlikely to match NVIDIA in AI chip capabilities, it notes that allowing exports of H200 AI chips to China would significantly expand available computing power for domestic AI firms, helping Chinese models narrow the gap with leading U.S. counterparts.

Easing AI chip export controls presents a major risk, the report warns. The report estimates that if NVIDIA exports three million H200 chips to China in 2026, it would provide at least a two-to-three-year boost in China’s AI computing capacity, and potentially longer. It adds that such volumes would give China more AI compute in 2026 than it could otherwise produce domestically until at least 2028 or 2029.

As a result, the report warns that large-scale H200 exports could enable China to build some of the world’s largest AI data centers. This would allow developers such as DeepSeek to close the gap with U.S. models much more quickly, particularly if computing power is concentrated in a limited number of locations.

Read more

• [News] China Reportedly Adds Huawei, Cambricon AI Chips to Procurement List Ahead of H200 Approval
• [Insights] Cambricon Remains China’s Top AI Chip Startup; Rumored 2026 Triple Output Faces SMIC Limits

(Photo credit: Huawei)