As MUFG America's February 2026 study shows , the most capable Huawei chip, the Attend 910C, is within spitting distance of Nvidia's H100 in compute power and is vastly more capable than the H20. It falls behind both in memory bandwidth, but not by egregious amounts. It's well behind the latest-generation Blackwell generation GPUs, but progress is clearly being made.
Huawei just announced its Atlas 350 AI accelerator based on its Ascend 950PR chip, promising almost three times the compute performance of Nvidia's H20. That could put it in the region of the H100 in terms of raw performance, leaving only Nvidia's Blackwell GPUs out in front, although a reported 1.4 TB/s memory bandwidth could represent a notable bottleneck.
Huawei has many more Ascend chips in the pipeline , but it's not the only one offering Nvidia competition. Alibaba unveiled its Zhenwu 810E AI chip in January , a chip said to be largely comparable to the H20. But it's 96GB of HBM2 memory delivers only around 700 GB/s of memory bandwidth; that's less than a quarter of the H20's.
Baidu announced its M100 and M300 AI chips in November last year, planning to launch them in 2026 and 2027, respectively. We couldn't find any direct Nvidia chip comparisons, but Baidu has suggested new supernode clusters of its chips could offer a 50% increase in performance over the last generation, suggesting major leaps in capability within short timeframes.
Cambricon's flagship Siyuan 590 AI accelerator falls behind Baidu and Huawei's efforts, yet it still expects to sell upwards of 500,000 units. The company is preparing its next-generation 690 chips for launch in 2026, although questions remain over whether it can get the materials it needs to manufacture them .
If Nvidia's AI hardware is its vanguard into the Chinese market, CUDA is supposed to be its backstop. That ecosystem is a protective moat that ensures those using CUDA-optimized software and Nvidia hardware keep doing so. But alongside Chinese hardware advances are Chinese software developments.
Baidu's Kunlunxin has translation layers that can run CUDA code efficiently, easing the transition from Nvidia hardware. Its PaddlePaddle framework is optimized for its Kunlun chips, too, so as Chinese providers make the switch, they can enjoy greater performance and efficiencies. Of course, there's also CANN, and this splintered approach could make China's local AI aspirations falter , if a unified approach is not fully adopted.
The longer Nvidia hardware isn't available, the more time Chinese companies and organizations have to transition to domestic alternatives that are easier to acquire, often come with government incentives, and are only likely to grow in power and efficiency, with an ever-more streamlined adoption pipeline.
That's what made it such a surprise that Nvidia told Tom's Hardware it wasn't planning to sell its Groq inferencing chips to China . ARM is very much looking to sell its new AI CPU there, though .
Nvidia's dominance is far from dead, and there really is no one likely to catch up in raw performance terms, or indeed in widespread support and compatibility for some time to come. But alternatives are rising, and the longer Nvidia's H200s sit in warehouses, with quiet production lines and strained trade routes, the more the tables will turn.
Jon Martindale is a contributing writer for Tom's Hardware. For the past 20 years, he's been writing about PC components, emerging technologies, and the latest software advances. His deep and broad journalistic experience gives him unique insights into the most exciting technology trends of today and tomorrow. ","collapsible":{"enabled":true,"maxHeight":250,"readMoreText":"Read more","readLessText":"Read less"}}), "https://slice.vanilla.futurecdn.net/13-4-19/js/authorBio.js"); } else { console.error('%c FTE ','background: #9306F9; color: #ffffff','no lazy slice hydration function available'); } Jon Martindale Freelance Writer Jon Martindale is a contributing writer for Tom's Hardware. For the past 20 years, he's been writing about PC components, emerging technologies, and the latest software advances. His deep and broad journalistic experience gives him unique insights into the most exciting technology trends of today and tomorrow.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/chinas-homegrown-silicon-suppliers-gain-traction-as-nvidia-struggles-to-get-its-chips-into-the-market-huawei-cambricon-and-more-step-up-to-fill-crucial-market-gap#main
- https://www.tomshardware.com
- Iran issues direct strike threat to Nvidia, Microsoft, Apple, Google, 14 other US tech companies — 'These companies should expect destruction of their facilitie
- Nvidia App adds 'Auto Shader Compilation' for faster load times in games — beta feature automatically recompiles shaders in the background after every driver up
- How Autonomous AI Agents Become Secure by Design With NVIDIA OpenShell
- Tryx Stage 360 AIO Review: Luxurious, unique, and quiet
- Kioxia discontinues 2D NAND products, last shipments to be made in 2028 — 1980s planar NAND memory reaches end of life
Informational only. No financial advice. Do your own research.