Nvidia's pared-down AI chips fight surging Huawei for China dominance.

In a high-stakes battle for dominance in China's burgeoning artificial intelligence chip market, US-based Nvidia is navigating a treacherous path, attempting to retain its foothold against the backdrop of stringent US export controls and the meteoric rise of domestic powerhouse Huawei. Nvidia's strategy hinges on a new generation of "budget" AI chips, specifically designed to comply with US regulations, yet the question looms large: can these pared-down offerings withstand the competitive onslaught from an increasingly capable Huawei, which enjoys strong domestic support and immunity from US sanctions within its home market? The answer will have profound implications for the global AI landscape and the future of semiconductor competition.

Nvidia's journey in the Chinese AI chip market has been a complex dance with US export regulations. Initial restrictions in 2022 barred the sale of its top-tier A100 and H100 AI processors to Chinese customers.[1][2] In response, Nvidia developed modified versions, the A800 and H800, to comply with those rules.[3] However, subsequent tightening of these controls in October 2023 and again more recently rendered even these chips non-compliant, forcing Nvidia back to the drawing board.[3][1][4] This led to the creation of a new trio of chips for China: the H20, L20, and L2.[3][5][6] These chips, particularly the H20, represent Nvidia's third attempt to create processors that meet US export parameters while still appealing to Chinese customers.[7] However, this compliance comes at a cost, with the H20 reportedly offering significantly less computing power than Nvidia's flagship international models and even appearing less powerful than Huawei's Ascend 910B in some key areas like FP32 performance, a crucial metric for common AI tasks.[8][9] Despite these limitations, Nvidia is projected to sell one million H20 chips in China in 2024.[10] The US government has signaled that even these downgraded chips, like the H20, will require licenses for export to China for the "indefinite future," citing national security concerns that they could be diverted to Chinese supercomputers.[2][11][12] This has already had a financial impact, with Nvidia reporting a projected $5.5 billion charge due to halted shipments and inventory reserves related to the H20 line.[1][11][12] To further navigate these restrictions, Nvidia is reportedly planning to release even more affordable AI chips for China based on its newer Blackwell architecture, but these too will feature reduced specifications, such as using GDDR7 memory instead of high-bandwidth memory (HBM) to stay within export limits.[13][14][15][16]

Meanwhile, Huawei has been aggressively capitalizing on the market disruption caused by US sanctions. Its Ascend line of AI chips, particularly the Ascend 910B, has emerged as a formidable competitor to Nvidia's offerings in China.[10][8] Huawei is reportedly ramping up production of the 910B and preparing to mass-ship its newer, more advanced 910C chip, with an even more powerful Ascend 910D in development, purportedly designed to rival Nvidia's previously unrestricted H100.[10][17][18][19][20] While the Ascend 910B might lag behind the H20 in certain aspects like inter-chip connectivity speed, it reportedly outperforms the H20 in raw processing power for some tasks.[10][8] More significantly, Huawei is not subject to US export restrictions for sales within China, giving it a crucial advantage in ensuring a stable and unrestricted supply for domestic customers. Chinese tech giants like Alibaba, Tencent, and Baidu, traditionally major Nvidia clients, are increasingly testing and adopting Huawei's Ascend chips.[21][22][23][24] ByteDance, for instance, reportedly ordered 100,000 Ascend 910B chips.[22][23] This shift is driven not only by the performance and availability of Huawei's chips but also by a push from Beijing to foster domestic semiconductor self-sufficiency.[17][9][25] However, Huawei faces its own challenges, including limitations in manufacturing capacity and yield rates due to restrictions on advanced lithography equipment.[10] Current estimates suggest Huawei sold around 200,000 Ascend 910B units in 2024, with plans to increase output.[10]

The battle for China's AI chip market extends beyond hardware specifications to the critical software ecosystem. Nvidia's CUDA platform has long been the industry standard, offering a mature and comprehensive suite of tools for AI developers.[26][17] This ecosystem has been a significant moat for Nvidia, making it difficult for competitors to dislodge its dominance. Huawei is actively working to build a similar ecosystem around its Compute Architecture for Neural Networks (CANN).[26][17][27][28] While CANN is still developing and lacks the widespread adoption and developer familiarity of CUDA, Huawei is investing heavily in its growth, including deploying engineering teams to major Chinese tech companies to help them migrate and optimize their code for the Ascend platform.[26][21][29] Some reports suggest that the Ascend 910C can achieve up to 60% of the H100's inference performance, with potential for more through CANN optimizations.[29][20][30] The willingness of Chinese developers to transition from the deeply entrenched CUDA to the newer CANN will be a critical factor in Huawei's long-term success.[26][17] For now, some Chinese firms are reportedly purchasing both Nvidia's H20 and Huawei's chips, balancing immediate needs with the strategic imperative to support domestic technology.[9] There are even efforts by companies like Moore Threads with its MUSA platform, which aims for CUDA code compatibility, to ease this transition.[31]

The strategic gamble for Nvidia is whether its compliant, lower-performance chips can maintain enough market share to remain relevant in China, a market that accounted for a significant portion of its revenue – around 14% to 20-25% of data center revenue in recent years.[7][32][33] While the H20 is priced competitively against Huawei's Ascend 910B, and in some cases even at a discount, the performance trade-offs are a significant consideration for Chinese AI firms striving for cutting-edge capabilities.[8][24] Nvidia's CEO, Jensen Huang, has acknowledged the competitive strength of Huawei and the risk of being excluded from the Chinese market if Nvidia cannot provide viable solutions.[34][24] The US restrictions, while intended to slow China's AI advancement, are inadvertently accelerating China's push for technological self-reliance, with Huawei at the forefront of this effort.[1][20][25][16] This situation could lead to a further fragmentation of the global AI chip market, with distinct ecosystems developing around US and Chinese technologies.[1][35]

In conclusion, Nvidia's strategy of offering budget, export-compliant AI chips for the Chinese market is a precarious balancing act. The company faces intense pressure from US regulations that limit the performance of its offerings, coupled with the rapid ascent of Huawei, which is leveraging strong domestic support and the imperative for technological independence to capture market share. While Nvidia's H20 and upcoming Blackwell-based China-specific chips may secure some sales due to the established CUDA ecosystem and immediate availability for certain applications, Huawei's improving Ascend series and its sanction-free status within China pose a formidable long-term challenge. The outcome of this rivalry will not only determine Nvidia's future in one of the world's largest AI markets but will also significantly shape the global AI industry's competitive dynamics and the trajectory of China's technological ambitions.