Jensen Huang's National Robotics Week Speech: I See the Turning Point of AI from "Conversation" to "Action"

Jensen Huang gave a speech at National Robotics Week last week, and I couldn’t calm down for a long time after watching it.

Honestly, my previous attention to NVIDIA mainly focused on GPUs and data centers — after all, training large models requires massive compute, and NVIDIA is an unavoidable part of this chain. But what Jensen Huang talked about this time made me realize NVIDIA’s ambitions go far beyond “selling shovels.”

His core point is this: AI’s next battlefield isn’t in the cloud, but in the physical world.

What does this mean?

The AI breakthroughs we’ve seen in the past few years have mainly focused on the “digital world” — ChatGPT can write articles, Midjourney can draw, Claude can write code. These are all digital content generation and understanding. But Jensen Huang believes AI’s next wave will penetrate the physical world — robotics, autonomous driving, industrial manufacturing, agriculture, and other fields.

He calls this direction “Physical AI.”

You might ask, what does this have to do with embodied intelligence and humanoid robots? It has everything to do with it.

The reason traditional robots have remained “clumsy” after decades of development is their poor generalization ability — factory robotic arms can only do fixed actions in fixed positions; change something slightly and they’re confused. But AI’s intervention gives robots “eyes” and “brains,” allowing them to understand the environment, understand tasks, and then make autonomous decisions.

This changes everything.

Jensen Huang mentioned a key data point in his speech: robot zero-shot learning success rates are rapidly increasing. This means robots no longer need to be reprogrammed for each new task, but can generalize through natural language instructions.

If this technology truly matures, what does it mean?

It means robots are no longer limited to working in factories; they can enter homes, hospitals, and restaurants. The reason the concept of “embodied intelligence” is so hot in 2026 is because everyone sees this possibility.

I previously discussed Unitree Technology’s IPO, with some saying the “mass production year” of humanoid robots has arrived. But honestly, I’m still conservative about this judgment.

There’s a huge gap between technology maturity and commercialization.

However, some details in Jensen Huang’s speech made me feel that this might really be accelerating:

First is the maturity of simulation platforms. Previously, training robots required repeated trial and error in real environments, with extremely high costs. Now you can train millions of times in virtual environments, then migrate to real robots. This greatly accelerates iteration speed.

Second is the evolution of edge hardware. Robots need real-time inference at the edge, placing high demands on chip compute and power consumption. NVIDIA’s Jetson series chips have made good progress in this area.

Third is ecosystem completeness. NVIDIA has gathered a bunch of partners, including domestic leading players like Zhiyuan Robotics and Unitree Technology. This “platform + ecosystem” approach is somewhat similar to how Android rose back in the day.

But the problems are also obvious: cost.

A humanoid robot capable of work currently costs anywhere from hundreds of thousands to over a million. When can it drop to tens of thousands to enter ordinary homes? This time window might be 5-10 years, or possibly longer.

And don’t forget, safety issues. After robots enter homes, what if they go out of control and hurt someone? These are social issues that need to be solved, not just technical ones.

My own judgment is: the “iPhone moment” of embodied intelligence will definitely come, but not in 2026, more likely in the 2030s. Technology needs time to mature, costs need time to drop, and social acceptance needs time to develop.

However, for investors and practitioners, now is indeed the time to enter — the earlier you position, the more advantage you have when technology matures. Just like mobile internet in 2010, AI entrepreneurship in 2015, and large models in 2020.

Dear readers, how long do you think it will take for embodied intelligence and humanoid robots to be deployed at scale? Is it the ultimate form of AI, or another overhyped concept?

Honestly, I don’t have a definitive answer to this question either. But one thing I’m certain of: whoever solves the problem of “making robots work stably in real environments” first will hold the key to the next wave of AI.

Jensen Huang obviously thinks so too, otherwise he wouldn’t be investing so heavily in the robotics field.

This battle will be worth watching.