Zhipu Runs Inference on 100,000 Domestic AI Chips, Claims 80% Cost Cut Without NVIDIA GPUs

Zhipu AI says it has moved inference workloads onto 100,000 domestically produced AI chips, a scale-up that it says can cut AI costs by 80% without relying on NVIDIA GPUs. The claim, reported by Chinese tech outlet MyDrivers, positions large-scale domestic chip deployment as a viable path for reducing the cost of running AI models.

The 100,000-chip milestone is significant because domestic Chinese AI accelerators have typically been viewed as a fallback constrained by ecosystem and toolchain gaps, not a primary production platform.

Zhipu is one of China's leading large-model developers, and its willingness to run inference at this scale on domestic hardware signals that the country's AI supply chain is maturing beyond the design phase, with domestic chips becoming part of real production environments rather than lab experiments.

A cost reduction of this magnitude, if realized in production, would change the economics of model serving for Chinese AI companies facing constrained access to advanced NVIDIA hardware, and could make large-model inference a routine operational expense for more enterprises.

The move also reflects a broader industry push to build AI infrastructure that does not depend on NVIDIA's roadmap, a trend that is accelerating across China's AI ecosystem, with scaled adoption by leading model vendors as a key step.

What to watch next is which domestic chip vendors, which models, and which deployment scenarios are involved, since the headline claim leaves open questions about real-world throughput, latency, and total cost of ownership, and those numbers will determine whether other companies can replicate the approach.