China builds its own 10,000‑card cluster

Alibaba has built a new kind of artificial intelligence factory: a 10,000-card computing cluster filled with its own Zhenwu chips instead of the Nvidia processors that have powered most big artificial intelligence systems for years. The system was launched with China Telecom and is designed for both model training and live inference work inside China. (scmp.com) (cnbc.com) A “card” in this context is a plug-in computing board built to do the heavy matrix math behind large language models, image generators, and recommendation engines. When companies talk about a 10,000-card cluster, they mean thousands of these accelerator boards wired together so one model can be split across many machines and run as a single pool of compute. (cnbc.com) (scmp.com) That scale matters because modern artificial intelligence models are too large for one server. Training a frontier model means chopping the work into thousands of parallel tasks, sending them across high-speed links, and keeping every chip fed with data fast enough that expensive hardware does not sit idle. (scmp.com) (arxiv.org) For most of the last few years, the default answer to that problem was Nvidia. Its graphics processing units became the standard not just because the chips were fast, but because the software stack, networking, and developer tools around them made it easier to build large clusters that actually worked in production. (cnbc.com) (scmp.com) China’s cloud companies have been trying to reduce that dependence for a simple reason: export controls have made access to top-end Western artificial intelligence chips less certain. When supply is restricted, the bottleneck shifts from “Who has the best model?” to “Who can still get enough hardware to serve one?” (scmp.com) (cnbc.com) Alibaba’s answer is vertical integration. The company already runs one of China’s biggest cloud businesses, develops the Qwen family of large language models, and has spent years designing chips through its semiconductor arm, which gives it a way to connect hardware, software, and cloud distribution inside one company. (alibabacloud.com) (scmp.com) That is what makes this announcement more than a chip story. A domestic processor on its own can still be a lab project; a 10,000-card cluster tied to a national telecom operator and a commercial cloud platform is much closer to a working industrial system that customers can actually rent and use. (cnbc.com) (scmp.com) The timing also matters. South China Morning Post reported that Alibaba’s launch came just after Shenzhen activated what it described as China’s first 10,000-card intelligent computing cluster built with Huawei Ascend 910C chips, which suggests this is turning into a race to prove domestic scale, not just domestic design. (scmp.com) In practical terms, these clusters are built for two jobs. Training is the long, expensive process of teaching a model from huge data sets, while inference is the cheaper but constant work of answering user prompts millions of times after the model is deployed. Alibaba and China Telecom said the new facility is meant to do both. (cnbc.com) (scmp.com) Inference is where the economics get especially interesting. A company can train a model once, but if that model becomes popular, it may need vast amounts of serving capacity every day, which is why a region that can assemble enough domestic chips and data centers can still build a meaningful artificial intelligence industry even without matching the very top Western training hardware. This is an inference based on how large model deployment works and on Alibaba’s statement that the cluster will support commercial cloud offerings and internal development. (cnbc.com) (blackridgeresearch.com) Alibaba’s own model strategy fits that picture. Its Qwen3 family, introduced in April 2025, spans models from small systems to very large ones, including dense models and mixture-of-experts designs, which are architectures meant to use only part of a model for each task instead of every parameter every time. That can lower the amount of compute needed for some workloads. (alibabacloud.com) (arxiv.org) The larger pattern is that artificial intelligence infrastructure is starting to split into regional stacks. In one stack, companies rely on Western chips, software, and cloud vendors; in another, companies assemble local chips, local clouds, local telecom partners, and local models tightly enough that the whole system can keep moving even when one global supply line closes. (scmp.com) (cnbc.com) Alibaba’s 10,000-card cluster does not prove that China’s domestic chips have caught up to Nvidia on every performance metric. It does show something narrower and more concrete: by April 8, 2026, Chinese companies were no longer talking only about prototypes and policy goals, but operating production-scale artificial intelligence infrastructure built around homegrown hardware. (cnbc.com) (scmp.com)