Beijing is pouring $295 billion into a national AI data center network, but no fully homegrown frontier training run has been completed yet
China has a $295 billion plan to stop relying on Nvidia. The country aims to build and interconnect a national network of AI data centers by 2028, with a mandate that at least 80% of the technology, including the AI chips doing the heavy lifting, comes from domestic suppliers like Huawei.
The plan, the money, and the mandate #
China plans to spend roughly 2 trillion yuan, approximately $295 billion, building out AI infrastructure over the next few years. That’s embedded in the country’s upcoming 15th Five-Year Plan period, which frames AI compute independence as a strategic national objective by 2028.
High-level directives from Chinese leadership are pushing for what officials describe as an “independent and controllable” AI ecosystem, aimed at ensuring that tightening US export controls can’t kneecap China’s AI ambitions.
Huawei sits at the center of this push. Analyst projections suggest the company could manufacture 3.3 million Ascend-series accelerators by 2028, potentially covering more than half of China’s domestic AI compute demand.
Progress, but not at the frontier #
Zhipu AI completed training of its multimodal model GLM-Image entirely on Huawei Ascend chips in January 2026, demonstrating that domestic hardware could handle a full training cycle for a competitive model.
Meituan claimed in June 2026 that it completed end-to-end training of its LongCat-2.0 model, a 1.6 trillion parameter system, on a domestic cluster of 50,000 chips.
Neither of these represents what the AI community considers a “frontier” pre-training run. The most critical AI models being developed in China still rely on Nvidia hardware or hybrid setups that mix domestic and foreign chips.
The telecom giants are all in #
China Mobile has laid out plans to triple its AI computing power by 2028 using exclusively homegrown chips. The target is reportedly 100 EFLOPS of AI computing capacity, up from levels at the end of 2024, and may include deploying a 100,000-GPU cluster built entirely with domestic accelerators.
The domestic AI accelerator market share is projected to hit approximately 56% by 2026, driven by US export restrictions creating supply gaps and Chinese policy actively favoring local suppliers through procurement requirements and subsidies.
What this means for the global chip race #
Huawei’s projected output of 3.3 million Ascend accelerators by 2028 would represent a supply base large enough to fundamentally change the economics of AI compute in China. Even if US restrictions were loosened, the policy momentum behind homegrown chips would likely continue.
Designing competitive AI accelerators requires not just fabrication capability but software ecosystem maturity, compiler optimization, and developer tooling that Nvidia has spent over a decade building with CUDA. China’s chip ecosystem is improving rapidly, but replicating that software moat is arguably harder than manufacturing the hardware itself.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our