Alibaba trials homegrown AI inference chip to curb Nvidia reliance as cloud revenue jumps 26%

Alibaba (NYSE: BABA) has quietly built a new AI chip aimed at inference work, and it's already in testing, according to people briefed on the project. Unlike an earlier Alibaba processor that was manufactured by Taiwan Semiconductor Manufacturing (NYSE: TSM), this one is being made by a Chinese fabricator - a sign that the company is pushing harder into domestically sourced hardware.

Short and blunt: the new chip is designed to be more flexible than Alibaba's previous silicon, with a goal of handling a wider range of inference tasks rather than a single, narrow workload. That matters because the AI market in China has been wrestling with limited access to some of the highest-end U.S. processors.

Nvidia (NASDAQ: NVDA) remains the go-to for large-scale AI training, but U.S. export controls have complicated the picture. Nvidia built a China-specific H20 model after wider export curbs in 2023, and that model has faced its own regulatory fits and starts - sales were effectively blocked earlier this year under the Trump administration before permission to resume was granted last month. The H20 is not on the same performance tier as the H100 or Nvidia's Blackwell chips, though it has been the main channel for Nvidia gear into Chinese cloud customers.

So what's changing? Alibaba's move looks like a bid to cut reliance on foreign-sourced accelerators for at least some inference workloads, and to keep capacity in-house if access to cutting-edge Nvidia silicon gets bumpy again. That doesn't immediately remove demand for Nvidia - the H20 remains available now - but a domestically produced, general-purpose inference chip could shift the long-term dynamics of procurement for cloud players and AI firms in China.

There's also a supply-chain angle. Taiwan Semiconductor Manufacturing (NYSE: TSM) was the foundry for Alibaba's earlier AI chip. Moving production to a mainland fab reduces cross-border manufacturing exposure and could change order flows for outside foundries if the design is adopted at scale.

For a quick corporate snapshot: Alibaba's cloud unit recently reported a 26% revenue increase in the April-June quarter, topping estimates and showing steady demand for cloud services - the very workloads that might run on this new chip. Alibaba is already one of Nvidia's biggest cloud customers in China, so any pivot in hardware sourcing has both commercial and strategic weight.

In short: Alibaba is betting on a more flexible, locally produced inference chip to hedge geopolitical and supply risks. That's likely to be greeted with interest by Chinese cloud customers and a watchful eye by global chip makers. Will it be a game-changer or just another piece on the board? The testing phase will tell.