Huawei unveils Ascend 950PR: new AI chip claimed to be faster than Nvidia H20 in China

Huawei announced that by the end of March 2026 it will begin shipping a new AI accelerator, the Ascend 950PR. The company claims that in a number of parameters it surpasses Nvidia H20 — a chip that Nvidia released specifically for the Chinese market under US export restrictions.

New Round of Competition

The history is important not just because of the hardware itself. H20 was a compromise product from Nvidia for China: powerful enough for local demand, but fitting within US restrictions, and for this reason it quickly became a notable option for local data centers. Against this backdrop, the launch of Ascend 950PR looks like an attempt by Huawei to close a window of opportunity and offer major Chinese customers its own alternative without dependence on supplies from the United States.

This is a segment where demand is growing faster than supply: models are becoming heavier, inference is becoming more expensive, and access to cutting-edge accelerators remains politically sensitive. Therefore, for the Chinese market it is important not just to get another GPU, but to build a stable local supply chain — from the compute chip to memory and server integration. This is important both for model training and for enterprise inference.

The launch of Ascend 950PR strikes at this nerve exactly.

What's at Stake

Huawei's main thesis is performance. According to the developers, Ascend 950PR surpasses Nvidia H20 almost threefold in FP4 computing. This is not a universal "faster at everything" metric, but a specific metric, particularly important for part of modern AI workloads where companies are aggressively reducing precision for the sake of speed and cost. But even in this form, the claim is strong: Huawei shows that it is already ready to compete not only on import substitution, but also on real efficiency.

• Nearly threefold advantage in FP4 according to Huawei's claims
• Own HBM memory without reliance on an external supplier
• Stronger position in the AI server market
• Reduced dependence on limited supplies of foreign GPUs

A separate emphasis is on proprietary HBM memory. In conditions of HBM scarcity, this is not just a technical detail but a strategic asset: without such memory, a modern AI accelerator loses a significant portion of its value. If Huawei can actually equip the new chip with its own HBM, the company gains an advantage on two levels at once — in cost of goods and in supply stability. For corporate clients, this may be no less important than the benchmarks themselves in the coming quarters and new tender cycles.

Why This Matters

For China, Ascend 950PR is a signal that local manufacturers are gradually transitioning from catch-up assembly mode to an offensive against specific product niches. If previously the main argument for Chinese solutions was affordability, now the focus shifts to a combination of performance, controlled supply chain, and political predictability. For cloud providers, government structures, and large corporations, this reduces the risk that critical AI infrastructure will run up against external restrictions.

But Nvidia still has an important advantage — the ecosystem. One strong chip is not enough if there is no mature software, optimization tools, libraries, and familiar workflows for developers around it. Therefore, the real test for Ascend 950PR will begin not at the moment of first shipments, but when customers try to massively migrate inference and model training to it.

If Huawei can simplify the migration, the effect of the release will be far wider than just a hardware announcement.

What It Means

The launch of Ascend 950PR shows that US sanctions have not stopped the Chinese AI chip race, but have accelerated the creation of its own alternatives. For the market, this is bad news for Nvidia's monopoly and good news — for companies that need a more predictable and locally available compute base for AI in the coming years.