China Aims for 300% AI Chip Output Increase by 2026 as Alibaba Unveils New Inference Processor

Alibaba, the Chinese technology giant, has developed a new artificial intelligence (AI) inference chip, currently undergoing testing, to significantly reduce its dependence on foreign technology, particularly from Nvidia. This strategic move aligns with China's broader national initiative to bolster domestic semiconductor production and achieve self-sufficiency amidst escalating U.S. export restrictions. The chip is designed for deploying trained AI models across a wide range of workloads, rather than the more computationally intensive task of model training.

"Alibaba built a new AI inference chip to cut reliance on Nvidia in China," stated Rohan Paul in a recent tweet, highlighting the core motivation behind the development. The new processor is notably more versatile than Alibaba's previous internal designs, enabling it to handle a broader array of inference tasks. Crucially, it is engineered for compatibility with Nvidia's existing software stack, allowing engineers to reuse established code and tools with minimal modifications, easing adoption.

The development comes as China faces stringent U.S. export controls, which have limited access to advanced Nvidia chips like the H100 and Blackwell series. While Nvidia's H20 chip was permitted for sale, Chinese authorities reportedly discouraged its purchase, further accelerating the push for local alternatives. Unlike earlier Alibaba AI processors fabricated by Taiwan Semiconductor Manufacturing Company (TSMC), the new chip is manufactured at a Chinese facility, ensuring an onshore supply chain despite potential process constraints.

Focusing on inference, which involves running trained models, allows the chip to tolerate older manufacturing nodes if sufficient memory and bandwidth are available. This contrasts with AI model training, which demands the fastest mathematical units and densest links to shorten multi-week computational runs. Industry experts, however, note that Chinese manufacturers still face challenges in producing chips comparable to the most advanced U.S. products, particularly for high-end training tasks.

Alibaba's initiative is part of a wider trend among Chinese firms, including Huawei with its Ascend line and companies like MetaX and Cambricon, to develop homegrown AI chips. This collective effort aims to fill the hardware gap and reduce reliance on foreign suppliers, with a reported national goal to triple the country's AI chip output by 2026. Alibaba, a major cloud-computing provider, plans to integrate these chips into its cloud infrastructure, offering AI processing power as a service and reinforcing its strategic investment of $53 billion over the next three years in AI and cloud development.