Inference-Optimized AI Ecosystem

As AI shifts from training (building models) to inference (running them), China's hardware disadvantage shrinks. Training requires bleeding-edge chips; inference can run efficiently on older or less powerful hardware. China has stockpiled millions of GPUs and developed domestic alternatives like Huawei's Ascend 910C.

DeepSeek and other Chinese labs have made breakthrough advances in inference efficiency — techniques like mixture-of-experts, speculative decoding, and aggressive quantization that let smaller, older chips serve large models. The irony: US export controls may have accelerated this optimization.

The implication: even if China never matches NVIDIA's latest training chips, it may not need to. If inference is where most AI value is created (serving models to users, not training them), China's efficiency-focused approach could be the right bet. The H20 chip that NVIDIA was allowed to sell to China is optimized for exactly this use case.