Chinese AI developer Moonshot AI is seeking additional Nvidia Blackwell chips to train its next-generation model, Kimi K4. The request marks a shift toward a system with more parameters than Kimi K3, which launched in late July with 2.8 trillion total parameters.

The chip demand underscores the ongoing capacity constraints facing domestic AI development under U.S. export restrictions. Moonshot initially trained K3 using Blackwell processors accessed through at least two Chinese cloud providers, stringing together multiple eight-chip servers across different data centers because no single operator had enough capacity. For inference, the company uses Nvidia H20 processors and currently recommends a minimum of 64 accelerators to host the model at scale.