AI Chips: Chinese Ambition Meets American Restrictions

The AI chip industry has witnessed rapid developments recently, most notably the entry of NVIDIA H20 and AMD MI308 chips into China. This occurred under complex political and commercial conditions imposed by U.S. restrictions, which led to special arrangements allowing sales to resume in exchange for a percentage of the revenue going to Washington.

.

The U.S. government permitted the sale of these two American chips to China in exchange for a 15% cut of sales revenue, after halting shipments last April citing national security concerns. This "conditional" arrangement reopened the door for AMD and NVIDIA in the largest AI market outside the U.S. However, it also increased the cost of ownership for Chinese customers and reshaped the balance of trade-offs with local alternatives, particularly from companies like Huawei.

A Technical Comparison of Processors

NVIDIA H20 Chip The NVIDIA H20 is built on a modified Hopper architecture to comply with Washington's export regulations. It features 96GB of HBM3 memory with a bandwidth of nearly 4TB/s and performance up to 296 teraflops at FP8 precision. While its performance is lower than the H100/H200 chips, its specifications make it highly effective for running large language models in inference workloads, especially for cloud services that require ultra-fast responses to millions of simultaneous user queries. This combination of high memory capacity and wide bandwidth makes it a preferred choice for developers in production environments that demand performance stability and scalability, even if it's not the ideal chip for advanced AI model training.

AMD MI308 Chip The AMD MI308 chip is based on the CDNA accelerated computing architecture and has been adjusted to meet U.S. export restrictions for China. It supports high-bandwidth HBM3 memory, which ensures stable performance in both training and inference tasks. Although its overall performance is lower than the flagship MI300 processors, it is designed to strike a balance between power efficiency and processing capability, making it a practical choice for medium to large AI workloads in cloud computing environments. The MI308 benefits from AMD's open-source ROCm ecosystem, giving it the flexibility to support popular software frameworks like PyTorch and TensorFlow, and facilitating its integration into cloud systems.

The Rise of Chinese Alternatives

The Chinese market landscape is incomplete without considering local alternatives, prominently led by Huawei's Ascend series, with the Ascend 910C/910D at the forefront. These chips are supported by the complex CloudMatrix system, which links hundreds of accelerators across data centers. The Huawei 910C processor is a strong competitor for inference tasks, offering higher memory capacity than the H20 in some configurations and deep integration with Huawei's software like MindSpore. This local approach has gained significant momentum, especially with the successful operation of the DeepSeek R1 model on arrays of 910C processors, demonstrating Huawei's strategy to achieve higher performance by networking more chips to compensate for the technical gap with NVIDIA.

Shipments of the Huawei 910C chip to major clients in China have expanded significantly over the past two quarters, with indications of orders from major internet companies like ByteDance and Baidu, as well as several Chinese telecom firms. This is redrawing the map of hardware adoption within local cloud platforms.

A Fierce Competition

The current situation in the Chinese market highlights new trade-offs. On one hand, the U.S. deal has brought back the H20 and MI308 chips, but it has also burdened the supply chain with a direct 15% financial cost, which analysts expect suppliers to partially pass on to customers. On the other hand, Chinese alternatives are now ready for large-scale inference and partial training, with the added benefits of local availability, faster delivery, and dedicated support.

A "Financial Times" report went further, noting that NVIDIA's market share in China has sharply declined over the past four years due to Huawei's rise, a shift largely accelerated by the U.S. controls themselves. In this context, Chinese companies managing sensitive large language models are opting for a hybrid strategy to mitigate risks. They rely on Ascend processors for inference layers where clustered arrays are highly effective, while reserving platforms with H20/MI308 processors for specific tasks that benefit from the CUDA/ROCm software ecosystems. This hybrid approach is likely to expand as long as the cost of regulatory compliance and the 15% revenue cut continue to weigh on the American offerings, driving Chinese companies to seek reliable local alternatives.