The 100,000-GPU Era: China Reaches Independence in AI Compute

On 10 July 2026, China's AI infrastructure reached a critical scale of 100,000 GPUs (十万卡). It got there despite US chip sanctions, becoming the first major economy to achieve full independence in AI compute.

From ban to independence

Five years ago, China's AI industry relied on American accelerators, above all NVIDIA GPUs. US sanctions, which began in 2020 and tightened in the following years, cut off supplies of high-performance processors. China, where 「万卡」 (10,000 GPUs) used to count as a success, had to build its own chips and software ecosystem.

• US chip sanctions began in 2020 and tightened by 2024–2025
• Previous milestone: 万卡 (10,000 GPUs or fewer)
• New milestone: 十万卡 (100,000 GPUs) in July 2026
• Investment in domestic chips by Huawei, Alibaba and Baidu

Three pillars of independence

The breakthrough rests on three pillars. First, in-house processors: Huawei, Alibaba and Baidu invested in chip design and had the chips built by contractors not covered by US restrictions. Performance lagged behind NVIDIA, but scaling up the architectures delivered results.

Second, software optimization. Instead of Western frameworks, Chinese teams built their own: PyTorch alternatives, compilers, and low-level optimizations for domestic silicon. In parallel, they trained research models on Chinese chips.

Third, local standards and integration. Rather than relying on foreign cloud platforms, they built domestic infrastructure that links companies in networks free of export-control dependence.

What does the turning point mean?

The 100,000-GPU figure is not arbitrary. At this scale the infrastructure becomes self-sufficient: there is enough performance to train giant language models (hundreds of billions of parameters), the software covers the entire development stack, and there are enough chips to scale across the country.

«This is not only a technical milestone, it is a political symbol: a

guarantee of independence from sanctions,» analysts in the research sector argue.

For developers, this frees their hands: there is no longer any need to wait on Western export controls. For the state, it is insurance against future restrictions: an attempt to cut off American chips would no longer derail its AI ambitions.

What this means

The geopolitics of AI is changing fundamentally. America's monopoly on compute, held for 10 years, is breaking up. Each major region will have its own AI infrastructure: America, Europe, China, possibly India. That will speed the global spread of models while fragmenting AI standards. The whole world, from small startups to states, will choose among several compute regions instead of one.

Frequently asked questions

What is 十万卡?

十万卡 (one hundred thousand GPUs) names a generation of AI infrastructure. Where people once talked about scale in 万卡 (ten thousand chips), the 十万卡 era means systems of one hundred thousand high-performance accelerators.

Why is this a historic achievement?

For the first time, a country has reached the volume of domestic chips and optimized software needed to be fully independent of Western export controls when developing large language models.