Why Nvidia’s H200 is unlikely to derail chip ambitions of China’s Huawei, Moore Threads

The US approval for Nvidia’s H200 artificial intelligence processors will give China a much-needed boost in computing power rather than pose a direct threat to domestic chips, even though the US chip is more powerful than rivals from Huawei Technologies and Moore Threads, according to analysts.\nWith a total processing performance (TPP) of 15,832, the H200 has surpassed all AI chips made by domestic vendors on the market so far as well as Nvidia’s H20, which was previously the most advanced Nvidia processor allowed to be shipped to China, according to a report from Bernstein Research. TPP measures how many calculations a chip can complete per second.\nThe H200, the latest generation of Nvidia’s Hopper-series chips, is an advanced version of the H100, which has been banned from export to China since 2022.\nThe H200’s export greenlight by the Trump administration is expected to provide a boost to China’s major cloud service providers, including Alibaba Group Holding, Tencent Holdings and ByteDance, which need the chips to build cloud infrastructure and support proprietary AI models. Alibaba owns the Post.\nDomestically, rivals to the Nvidia chip included processors made by Huawei and Alibaba’s chip design unit T-Head, according to the Bernstein report.\n\nHuawei’s Ascend 910C, which reportedly started production earlier this year, has a TPP of 12,800, while Alibaba’s PPU 2.0 is estimated to reach a similar level, though its production status remains unclear.\nAnalysts did not see the H200 as an immediate threat to domestic vendors, as China’s huge demand for AI computing power would provide growth opportunities for both US and domestic chips. Nvidia’s chips were expected to remain the top choice for AI training among big tech companies.\nThe impact of H200 sales would be “minimal” on domestic graphics processing units (GPUs), mostly used in inference tasks that required less performance compared with the more demanding training tasks, according to semiconductor industry analyst Zhang Haijun.\nThe export approval meant that Chinese clients would have access to a “more advanced and mature computing power option”, while domestic chip platforms would get more “development opportunities” when adopted alongside H200 chips, said Zhao Liangbi, an analyst at China Galaxy Securities.\nBernstein also expected home-grown chips to continue to increase their market share in the coming years.\n“Currently, the biggest bottleneck lies in constrained local advanced logic production capacity, which has limited the output of local AI chips,” Bernstein analysts said, adding it anticipated the supply of local AI chips to grow five times in the next three years as domestic companies ramp up their production capacity.\nChinese chipmakers are nevertheless doubling down on efforts to catch up with Nvidia amid China’s broader efforts to cut reliance on US technologies.\nIn September, Huawei unveiled a three-year road map for its Ascend AI processors, including Ascend 950, 960 and 970, signalling its ambition to play a central role in China’s tech self-sufficiency efforts.\nMoore Threads, the Beijing-based chipmaker dubbed as “China’s little Nvidia”, recently raked in 8 billion yuan (US$1.1 billion) from its initial public offering in Shanghai – the mainland’s second-largest stock sale this year – and was expected to accelerate the company’s GPU ecosystem development.\nDespite the ambition, Moore Threads’ latest processor MTT S400 was generations behind the H200, with a TTP of 1,568, just one-tenth of Nvidia’s chip, according to Bernstein.\nFor Nvidia, its “moat” among Chinese developers remains the CUDA software ecosystem, as many AI systems still rely on that stack.\nWhile Huawei is also attempting to standardise its own instruction set, the open-source CANN, Nvidia’s chips are still preferred by companies such as DeepSeek for training advanced AI models.\n“At this point, none of the [domestic] systems have proved to be more stable than CUDA, so Nvidia cards will remain my first choice,” said Richard Xu, who develops large language models at a research institute in Shanghai.\n