Last year at CES, Jensen Huang took the stage wearing his signature leather jacket and told a grand story — the era of physical AI is here. Blackwell chips, million-token long context, agent processing, autonomous driving decisions… all sound promising.

But this time is different. This year, NVIDIA is seriously tackling a real problem: how to truly reduce the cost of inference.

This is crucial. Because since open-source inference models like DeepSeek R1 became popular, the entire industry has realized — when collaboration is truly unleashed, technology diffusion happens at an incredible speed. Although open-source models are still about half a year slower, every six months they can catch up, with downloads and usage skyrocketing.

Reality is forcing NVIDIA to change its strategy. Having chips alone is not enough; the entire infrastructure for inference — computing power, networks, storage — must be completed to lower inference costs from the source. This isn’t just about training scores for super-large models, but about truly embedding AI capabilities into real-world scenarios like autonomous driving and robotics.

This means the industry landscape is quietly changing. The significant reduction in inference costs will directly impact hardware demand, and the level of infrastructure development will become a key factor in determining who can deploy applications quickly.