Alibaba exports a non‑NVIDIA stack
Alibaba has launched a data centre built on its own chips and is running DeepSeek models on Huawei systems, signalling an emerging AI hardware stack that does not depend on NVIDIA. That development illustrates how large cloud and integrator customers are starting to field alternative stacks for training and inference (x.com).
For years, building a top artificial intelligence system meant buying NVIDIA chips first and figuring out the rest later. This week’s surprise is that Alibaba opened a new data center in southern China with 10,000 of its own Zhenwu processors, while DeepSeek’s next big model is being prepared for Huawei’s Ascend chips instead. (cnbc.com) (usnews.com) A data center is a warehouse full of computers, and an artificial intelligence chip is the engine inside those computers that does the heavy math. If one company designs the engine, another writes the software, and a third runs the cloud, the whole machine can stall when one part is cut off. (cnbc.com) (cset.georgetown.edu) NVIDIA became the default because its chips and its software tools arrived together. That pairing made it easier for model builders to train giant systems and then serve answers to users at scale. (cset.georgetown.edu) (cnbc.com) The break came from Washington before it came from Beijing. The United States tightened advanced chip export controls in October 2022 and expanded them again on October 17, 2023, specifically targeting the kind of computing used for artificial intelligence training. (csis.org) (cset.georgetown.edu) That pressure pushed Chinese cloud companies to build more of the stack themselves. Alibaba already had the cloud business, the Qwen model family, and a chip unit called T-Head, so it could start linking silicon, software, and customers inside one system. (cnbc.com) (alibabacloud.com) (github.com) The new Alibaba facility was built with China Telecom in Shaoguan, Guangdong, and Alibaba says it can handle both training and inference. Training is the expensive part where a model learns patterns from huge data sets, while inference is the cheaper part where the trained model answers your prompt. (cnbc.com) (finance.yahoo.com) DeepSeek sits on the other side of the same shift. Reuters reported on April 3 that DeepSeek’s V4 model is expected to run on Huawei hardware, after work to adapt parts of the model to Huawei’s Ascend platform. (usnews.com) Huawei matters here because a chip is not enough by itself. A model builder also needs the networking, system software, compilers, and tuning work that let thousands of chips act like one giant computer, and Huawei has been pushing that full package with Ascend. (bis.gov) (cnbc.com) Reuters also reported in late March that Alibaba and ByteDance were planning orders for Huawei’s Ascend 910C chips. That means the same company can be a customer of Huawei in one part of the market and a supplier of its own silicon in another, which is what a real hardware ecosystem looks like when it starts to form. (cnbc.com) None of this means NVIDIA is gone. It means the market is changing from one dominant stack to several regional stacks, with China’s biggest cloud groups now proving they can train models, run services, and sell compute without waiting for NVIDIA to be available first. (cnbc.com) (usnews.com)
