Chinese AI Giants Betting Big on Custom Silicon

Is the era of relying solely on NVIDIA coming to an end for Chinese AI labs? Recent reports suggest that major players like Zhipu AI and DeepSeek are actively exploring or even developing their own custom AI chips.

What I find particularly interesting here is the strategic shift in focus. While everyone used to obsess over training capabilities, DeepSeek seems to be targeting the "inference" stage. This makes perfect sense—as these models go mainstream, the cost of running them (inference) becomes a much bigger bottleneck than the initial training phase. By designing chips specifically for inference, they can achieve much higher efficiency than using a "one-size-fits-all" GPU.

We are seeing a massive trend of vertical integration. It's not just China; OpenAI, Anthropic, and even SpaceX are moving toward custom silicon. For companies like Zhipu, this isn't just about escaping US export controls or reducing reliance on Huawei; it's about "software-hardware co-optimization." When you own the architecture, you can squeeze every bit of performance out of the silicon specifically for your model's weights.

It's a high-risk, high-reward play. Designing chips is incredibly expensive and capital-intensive, but if they pull it off, they won't just be AI companies—they'll be self-sustaining tech ecosystems.

What do you guys think? Will custom inference chips become the new standard for LLM providers, or is it too much overhead for a software company?