特朗普政府对向中国销售 AI 加速器的出口管制政策并未阻止华为自研产品填补空缺。本周，华为轮值董事长徐锐在一份问答式声明中声称，其 Ascend 神经处理单元（NPU）在中国市场份额已超过英伟达。“很难收集英伟达在中国的市场份额数据，但根据我们收集的数据，Ascend 已经超过英伟达，”他表示。“中国人有强烈的紧迫感，不会接受别人决定我们是否能获得某些产品的未来。”
The Trump administration's export control policies governing the sale of AI accelerators to China haven't stopped homegrown Huawei from filling the void. In a Q&A-style release this week, the company's rotating chairman Eric Xu claimed that its Ascend line of neural processing units (NPUs) had exceeded Nvidia in Chinese market share. "It's pretty hard to collect data about the market share of Nvidia in China, but based on the data we have collected, Ascend has surpassed Nvidia," he said. "Chinese people have a keen sense of urgency, and will not accept a future in which others decide whether or not we can have access to certain products."
自上任以来，特朗普政府在出口管制上彻底转变。2025 年 4 月对英伟达和 AMD 的中国规格加速器实施新许可要求后，商务部在夏季反转政策，随后达成协议，让美国政府在某些面向中国的芯片销售中获得一定比例的收益。到去年年底，特朗普政府进一步放宽限制，宣布首次允许英伟达向获批的中国客户出售更强大的 H200 系列芯片。
Since taking office, the Trump administration has done an about-face on export controls. After imposing new licensing requirements on Nvidia's and AMD's China-spec accelerators in April 2025, the US Commerce Department reversed course that summer, later reaching an arrangement to cut the US government in on a percentage of certain China-bound chip sales. By late last year, the Trump administration loosened restrictions further, announcing that Nvidia would be allowed to sell its more potent H200-series parts to approved Chinese customers for the first time ever.
然而，损失已经造成。虽然美国政府已批准黄仁勋恢复销售，但年初对 H20 的初步封锁已在北京引发警报。当地政府官员开始施压比特农场运营商，转向国产替代方案。在英伟达第二季度财报电话会议上，管理层透露，虽然公司已向该地区发运少量 H200 加速器，但其占数据中心收入不到 1%。“即使我们的芯片技术不够先进，至少供应得到保障，这样你就不必担心芯片供应问题，”徐锐说。
However, the damage was already done. While Uncle Sam had given Jensen Huang its blessing to resume sales, the initial blockade of H20s earlier in the year had set off alarm bells in Beijing. Government officials there began pressuring bit barn operators to transition away from foreign accelerators in favor of homegrown alternatives. During Nvidia's Q2 earnings call, executives revealed that while the company had begun shipments of a small volume of H200 accelerators to the region, they amounted to less than 1 percent of its datacenter revenues. "Even though our chips may be less advanced, at least their supply is assured, so that you don't have to worry about chip supply day in and day out," Xu said.
对于中国政府、国内产业以及华为来说，未来的道路无疑是推动芯片和整个半导体价值链的全面自给自足。
"For the Chinese government, for the domestic industry, and for Huawei, the path forward is undoubtedly to push for full self-sufficiency in terms of chips and the entire semiconductor value chain."
正如我们最近讨论的那样，虽然华为最新的加速器在纸面上仍落后于 Nvidia 或 AMD，但它们通过规模弥补了这一缺口。
As we recently discussed, while Huawei's latest accelerators still fall short of Nvidia's or AMD's on paper, they make up for that deficit with scale.
华为目前正在部署 256,000 卡 Atlas 950 超级集群，而其更新的架构旨在支持多万亿参数模型的训练和推理，并可扩展到多达一百万个 NPU。
Huawei is currently deploying a 256,000-card Atlas 950 SuperCluster, while its newer architecture is designed to support training and inference for multi-trillion-parameter models and scale to as many as one million NPUs.
然而，正如徐所指出的，华为在中国之外的市场上在一段时间内不太可能对 Nvidia 构成威胁。
However, as Xu notes, Huawei likely won't be a threat to Nvidia outside of China for a while.
“我们没有足够的产能满足中国的需求，也没有计划以全面的方式扩大国际市场，”他表示，淡化了华为已开始向友好国家供应其芯片的报道。
"We don't have enough capacity to satisfy the demand in China.
当被问及使用 Ascend 加速器为马来西亚主权 AI 计划提供动力时，他说：“满足国内客户与国际客户需求之间的平衡实际上相当容易。”
We don't have plans to expand the international market in a fully-fledged way," he said, downplaying reports that Huawei had begun supplying friendly nations with its chips.
“我们优先满足急需计算资源的中国客户；在中国之外，我们只向极少数客户提供服务，这些客户要么无法获得替代方案，要么想要替代方案。”
"The balance between satisfying the needs of domestic customers and those of international customers is actually quite easy," he said when asked about the use of Ascend accelerators to power Malaysia's sovereign AI initiative.
随着政府官员推动摆脱外国芯片的压力，以及软件范式的变化，使得模型开发者（如 DeepSeek）更容易过渡，DeepSeek 已与华为芯片合作超过一年。
"We prioritize Chinese customers who are in urgent demand for compute; outside of China we only supply a very limited number of customers who either have no access to alternative solutions or want an alternative solution."
“Ascend 首先遇到的障碍实际上并非硬件本身，”华为海思首席科学家廖恒说。
Along with pressure from government officials to transition away from foreign chips, changing software paradigms are making the transition easier for model devs, like DeepSeek, which has been working with Huawei's chips for well over a year now.
“主要是生态系统障碍，因为两年前 Ascend 仍面临巨大的软件障碍。”
"The first barrier encountered by Ascend is actually not hardware itself," said Liao Heng, chief scientist of Huawei's HiSilicon.
廖指的是通常所说的 CUDA 巢穴，这是一种软件障碍，并非中国独有，而是几乎所有竞争芯片制造商都面临的。
"It's primarily the ecosystem barrier, because two years ago there was still a tremendous software barrier for Ascend."
“过去18个月里发生了剧烈变化，”廖说。他解释说，二年前，AI训练主要由运行在CUDA之上的PyTorch主导，但模型架构的变化促使模型开发者转向“mega‑kernel‑style编程”。“CUDA不再被前沿实验室视为重要，远不如两年前那样重要，”他说。为此，DeepSeek据报道本周发布了几款软件工具，旨在让华为的Ascend加速器更易于其他AI实验室使用。®
Liao is referring to what's often called the CUDA moat, a software hurdle not unique to China but faced by essentially every competing chipmaker. "There have been drastic changes over the past 18 months," Liao said. He explained that, two years ago, AI training was dominated by PyTorch running atop CUDA, but changes in model architectures have pushed model devs to "mega-kernel-style programming." "CUDA was no longer deemed important by the frontier labs, nowhere near as much as two years ago," he said. To this end, DeepSeek reportedly released several software tools this week aimed at making Huawei’s Ascend accelerators more accessible to other AI labs. ®