Nvidia Should Not Be Threatened by Huawei’s Newest AI Chips

With U.S. sanctions sharply restricting the sale of advanced GPUs such as the H100 and H20 into China by Nvidia Corp. (NASDAQ: NVDA), questions naturally may arise about whether Chinese semiconductor chip manufacturer Huawei is positioning its new Ascend series as the go-to alternative in that country.

Huawei’s recent announcement of its new artificial intelligence (AI) training processors, the Ascend 910C and Ascend 910D, has generated excitement within China’s domestic technology community. In one sense the excitement is justified, since Huawei’s achievements are impressive.

That is more the case since Huawei does not have access to the ASML extreme ultraviolet EUV photolithography machines necessary for the most elite “3nm” and now “2nm” semiconductor process nodes. Until recently, it was believed EUV was necessary even for 7nm and 5nm.

Then SMIC (Semiconducutor Manufacturing International) (HKG: 0981), China’s largest chip manufacturer, came out with versions of 7nm and 5nm chips using only the legacy ASML DUV (deep ultraviolet) machines for use in Huawei smart phones. As we explained in Gilder’s Technology Report, SMICs versions were costly to make, and had lower yields and inferior performance compared to 5nm and 7nm chips made by Taiwan Semiconductor and other Western firms. Still, considering the handicaps overcome, it was an impressive performance.

Huawei’s AI chips tell a similar story: very impressive but not yet very relevant. The new Huawei semiconductor chips represent only a partial and mostly regional solution, not a serious global competitor to Nvidia’s dominance in AI training hardware.

Head-to-Head Comparison: Huawei Ascend 910C vs Nvidia A100

The Ascend 910C was Huawei’s first major step toward closing its gap with Nvidia. When benchmarked against Nvidia’s older A100 GPU, the 910C presents interesting technical specifications on paper. But deeper examination shows that Huawei’s architectural, software, and manufacturing limitations continue to leave it well behind in real-world AI training capabilities.

While the 910C matches or even exceeds the A100 on pure floating-point performance, its lower memory bandwidth, inferior interconnect, and weaker software stack limit its usefulness for cutting-edge AI models. In practice, the A100 remains more versatile and performant across a wider range of workloads.

Head-to-Head Comparison: Huawei Ascend 910D vs Nvidia H100

Huawei’s next major move is the Ascend 910D, which is positioned to rival Nvidia’s flagship H100 GPU. This chip represents Huawei’s highest achievement to date in AI processor design. Nevertheless, the gap between the 910D and the H100 remains large, especially when real-world deployment factors are considered.

Although the 910D closes some theoretical performance gaps on paper, it still falls short in the critical areas that matter for deploying large-scale, real-world AI training clusters. Memory bandwidth and interconnect performance are essential for large models, and Huawei’s solutions do not match Nvidia’s integrated systems yet.

Comparison of Nvidia H100 and H20

Given that Nvidia introduced the H20 specifically to comply with U.S. export controls to China, it is important to understand the key differences between the H100 and H20, particularly when evaluating what Huawei is really competing against inside the Chinese market.

The H20 was a downgraded version of the H100, significantly less powerful, but acceptable for training medium-sized AI models. Huawei’s Ascend 910C and 910D are effectively competing more against the H20 in the Chinese market rather than against the true H100. Even then, Huawei struggles to match the overall system-level performance Nvidia offered.

Shipment Volumes of Nvidia Chips to China in 2024

Another important context is the volume of Nvidia chips that made it to China before the April 2025 export tightening.

The H20 was the primary Nvidia product supporting China’s AI efforts in 2024, not the full H100. Huawei’s new chips are thus not facing the full brunt of Nvidia’s latest technology, but a compromise version designed for restricted markets.

Conclusion: Huawei’s Ascend 910C and 910D Chips Gain Ground for Geopolitical Reasons

Huawei’s Ascend 910C and 910D chips remain well behind Nvidia’s best offerings when evaluating critical factors like software support, interconnect scalability, memory bandwidth, manufacturing process quality, and power efficiency. Even when Huawei can match or exceed Nvidia’s older A100 or the downgraded H20 on theoretical metrics, it fails to replicate the integrated system-level advantages that Nvidia provides, particularly for training large language models or for managing massive AI data centers.

For geopolitical reasons, Huawei’s chips are gaining traction within China. With U.S. export restrictions limiting access to Nvidia’s H100 and H20 products, Chinese companies have little choice but to adopt domestic alternatives like the Ascend 910C and 910D. This momentum is driven not by technical superiority, but by necessity and national policy to achieve technological self-sufficiency. In this constrained environment, Huawei is benefiting from an artificial competitive advantage that would not exist under free-market conditions.