Jensen Huang: “We went from 95% market share to 0% in China.”
China and the U.S.A are locked in a titanic battle over tariffs.
The U.S. makes the world’s best AI Accelerators: Nvidia, AMD, Google, AWS …among others. But the U.S. worries China could deploy these for military purposes, so it imposed strict export controls that resulted in China getting the second-best AI accelerators. These export controls have been further tightened as part of tariff negotiations, cutting off GPU sales, at least for now. Case in point: Bytedance canceled its contract with Broadcom for an AI accelerator.
In August of this year, the Chinese Government mandated 50% of chips used in data centers be of Chinese origin. It seems China decided they’re better off learning to build their own AI chips. This is unlikely to change even if/when the U.S. relents on export licenses for GPUs.
Nvidia’s market share in China has dropped to 0%. Jensen Huang recently said this is costing Nvidia $35 billion or more per year. Even when exports of GPUs resume, likely 50% will be the limit for U.S. GPU sales. So who will fill the vacuum for AI in China, and can they match Nvidia?
Fig. 1: A Great Wall is rising between Chinese and U.S. AI hardware ecosystems. Source: Creative Commons
My conclusion is Huawei is best positioned; Alibaba and Baidu are next. It will be hard for them to match Nvidia anytime soon due to limitations of China foundry, advanced packaging, and HBM. But they are good enough for GenAI to grow in China. Let’s dig into the basis for this.
China is very capable, but is not a leader in AI semiconductors
China graduates more EE PhDs every year than the U.S., not counting those who study in the U.S. (and many return to China now). China’s universities rank only second to U.S. universities in the quality of EE graduate studies.
China is the world’s second largest economy. China is a leader in many aspects of high technology and controls more than half of the market for many critical technologies such as EVs, batteries, solar.
China has a fast-growing electric power infrastructure. China’s electric power is less expensive because China has the largest installed bases of solar, hydro, wind, and extensive nuclear power investment. And it will use coal if needed. The New York Times recently described the network of power lines that link massive solar and wind installations in the uncrowded Northwest of China to the densely populated East: China has more than 40 power lines, as much as 2,000 miles long, that carry more electricity than any transmission line in the U.S..
But in GenAI semiconductors China is not a major player. China foundries and chip companies lag the U.S., especially in advanced process nodes, advanced packaging, and GenAI accelerators. To catch up on advanced nodes will require fab equipment like ASML steppers, which are subject to U.S. pressure to restrain advanced exports. And HBM, a critical ingredient of all AI accelerators, is not made in China (yet).
China has some strong GenAI models such as Deepseek and Qwen (Alibaba). They lag OpenAI and Anthropic, but they are not far behind. There is demand for GenAI if China semis can supply it.
What are the major Chinese data center players, and what are they using?
China’s estimated total AI data center capital expenditure is almost $100 billion in 2025 (source: South China Morning Post). This is about 20% to 25% of the AI CapEx in the U.S. (Morgan Stanley estimates the U.S. big 6 hyperscalers are investing $430 billion in 2025).
The biggest China cloud service providers, according to Data Center Dynamics, are Alibaba (33%), Huawei (18%), Tencent (10%)
The major U.S. cloud providers have little or no presence in China:
Below are details of some recent China Data Center investments and AI chips selected (as a comparison, AWS’ datacenter for Anthropic is an $11 billion project with 500,000 Trainium-2s running now, with plan to expand to 1,000,000 Trainium-2 in 2026):
Alibaba Cloud
Huawei Cloud
Tencent Cloud
ByteDance Cloud
Baidu Cloud
China Mobile
China Unicom
Huawei shows up in most data centers; a couple use Alibaba and Baidu.
Huawei appears to be the clear No. 1 choice for replacing Nvidia. In 2025, Ascend production is expected to be 400,000 units with plans for over 1 million units in 2026. The latest Ascend is about 2/3 the inference performance of a Nvidia H100 (Source: Data Center Dynamics). Huawei used TSMC but has shifted to SMIC 7nm. A major issue is that HBM comes from Korea, which is subject to export controls. Huawei has stockpiled HBM for now, but at some point in 2026 export controls need to be eased to maintain production rates. Huawei appears to have developed a custom HBM, probably using China’s ChangXin Memory Technology (CXMT).
The Cloud Matrix 384 Cluster using Ascend 910C outperforms Nvidia GB200 NVL72 by 2x (at higher cost) because of its all-to-all topology, fully-optical (but not CPO) scale-up network (Source: SemiAnalysis). The interconnect enables a larger pod size than NVL72 that, for large models, more than offsets the lower performance of the Ascend GPU. Power consumption for Cloud Matrix is 4X the GB200 NVL72 for the same level of throughput. But in China, power is less expensive and readily available. So Cloud Matrix 384 can enable China data centers to run large model GenAI at scale. Cloud Matrix 384 won’t compete with Nvidia in the U.S. and Europe, where power is constrained, but it could in Saudi Arabia and the UAE where power is plentiful if GPUs are in short supply.
At Huawei Connect September 2025, Huawei outlined an aggressive roadmap, acknowledging that China foundries will lag on advanced process nodes for the foreseeable future. Their strategy is to create a new computing architecture based on SuperPod interconnect. They announced three new series of Ascend chips over the next three years that will double performance annually and improve support for low-precision data formats to improve throughput/watt. They are also increasing their interconnect bandwidth by 2X or more per year.
These new chips will be used with the Atlas SuperPOD which can accommodate 8,192 Ascend chips linked with all-optical interconnect. This will be available by the end of next year. Larger pods with low latency interconnects (Huawei has not specified theirs) dramatically improve performance for training and for the largest reasoning/agentic models. The next step is the Atlas Supercluster with over 500,000 AI accelerators. Huawei is thinking very big.
They also announced a shift in the future to their own proprietary HBM chips, which are lower-cost and more optimized for their needs. The manufacturer appears to be domestic (CXMT or SMIC would seem the only options).
Huawei doesn’t talk about how their roadmap compares to Nvidia on throughput/watt. They are making improvements in their architecture, but having to use trailing edge process nodes will likely mean they can do okay on throughput/$, but not so well on throughput/watt. Of course, in China throughput/watt is less of a problem.
Alibaba’s T-Head appears to be a distant second choice for replacing Nvidia. It is reported to be similar performance to Nvidia’s H20 (South China Morning Post) using HBM2e. It is built on SMIC 7nm using a chip-to-chip interconnect bandwidth, claimed at 700GB/sec in their Panjiu 128 super-node rack with 128 AI chips, which was announced at the Yunqi Conference. Using this and the HPN 8.0 high performance scale-out network, Alibaba can connect up to 100,000 GPUs for training.
Baidu’s third generation 7nm Kunlun is optimized for training, with Baidu deploying a 30,000-chip cluster that it claims can train 100 billion+ parameter models. The accelerators are connected using their XPU Link interconnect.
There are other options. Tencent is the other Cloud Service Provider (CSP) with its own inference chip — Zixiao. Others are non-CSPs, including Cambricon, Iluvatar, Enflame, MetaX, Moore Threads, and Biren.
The ultimate winners in China may be determined by politics as well as market forces.
China’s Achilles heel is that they do not have access to TSMC, which builds all of the GPUs/XPUs for Nvidia/AMD/AWS/Google. Nor do they have access to the leading HBM suppliers that support the GPUs/XPUs.
China’s only significant foundry on the global level is SMIC, with 5% market share in Q2/2025 (Counterpoint). TSMC dominates at 71%, Samsung 8%, UMC 5%, Global 4%, others 8%.
SMIC’s most advanced node is 7nm compared to 2nm for TSMC; 5nm is in development for 2026. SMIC is a couple generations behind, partly because it’s one-tenth the size and cannot do the same level of R&D, but also because it is deprived of the ASML steppers required for advanced nodes. Even to get to 7nm it has to do expensive and slow multi-patterning using the earlier generation steppers they do have. There are companies in China building steppers, but none that are close to ASML’s advanced-node capability. Further, SMIC’s advanced packaging capability and capacity lags far behind TSMC’s. SMIC may be challenged in meeting the demand for China accelerator companies since their share was <5% until recently — and some of that was made at TSMC and Samsung, which is no longer possible with U.S. export restrictions.
There are other foundries in China, but not with advanced nodes. Hua Hong’s most advanced node is 28/22nm.
DRAM is just as critical for GenAI accelerators. It needs to be in the form of HBM to enable on-interposer coupling of memory and accelerator. The leading DRAM company in China today is ChangXin Memory (CXMT), with HBM2 in mass production and HBM3 planned for 2026 and HBM3E in 2027. It lags Samsung/SKHynix/Micron, but HBM2/3 is sufficient to build reasonably competitive AI accelerators. Wuhan Xinxin (XMC) is a NOR/Flash player that is working on bringing up HBM capability.
China’s AI accelerators cannot compete with Nvidia because their foundry and HBM suppliers are one to two generations behind. But SMIC and CMXT are sufficiently capable to enable GenAI chips that can fill the vacuum and support rapid growth in GenAI – at higher cost and higher power. China has tremendous capability, so over time they may develop domestic technologies to enable them to compete, but if they do it won’t be until the 2030s.
Good article on a hard to research topic. One tough challenge – keeping current. I believe NVIDIA has improved inference on GB200 NVL 72 by about 4x since performance numbers vs Cloud Matrix 384 came out, thanks to software and system optimizations. Does that mean a single rack GB200 NVL is 2x faster than a 16 rack Cloud Matrix 384 ? Probably, and at 10x lower power !