Humanoid robots

With technologies that are slow to reach widespread adoption, there's always a risk that fatigue sets in. You see this with self-driving cars. We were promised self-driving cars for ages, but they kept getting pushed back a year or two. People felt they were being overhyped. Now we actually have self-driving cars and ironically the pendulum has swung back the other way, and if anything they're under-hyped.

With humanoids, I think you see something similar - people have been excited about the progress of the Boston Dynamics robots for long enough that they maybe underrate the potential of newer generations of humanoids.

• Humanoids have been massively decreasing in cost in recent years. Unitree made huge waves last year when they released the G1 for just $16000! It's only four feet tall, but this is still an insane achievement. Consider that at the time this was released, Elon Musk (notoriously a very optimistic estimator) was forecasting that the Tesla Optimus humanoid would retail between $20k and $30k once they scaled up production. Unitree haven't even begun scaling production yet, and they're in advance of Tesla's cheapest estimate! The result of this is that their robots are already widely used in industry and every Nvidia robotics video I've seen uses Unitree humanoids for testing. A new US startup k-scale labs is going even cheaper and looking to ship $9k humanoids with off-the-shelf parts and open-source firmware.

• Humanoid providers have aligned on a form factor. For a while in humanoids the only game in town was Boston Dynamics and while it was impressive to see their humanoids learn to backflip, their previous generations were fairly limited due to being hydraulic. BD ditched hydraulic Atlas for an all-electric model in April 2024 and last year we saw the leaders in humanoids (Tesla and Figure) all adopt a broadly similar form factor.

• Competition to deploy to production is accelerating. Already there's a heated race in progress between the leading humanoid providers: it doesn't quite have the NBA-level competition of frontier AI labs yet but it's moving in that direction. Both Tesla and Figure have begun making their use of their humanoids in production - probably at the moment this is largely for show, but the main thing is that the competitive dynamics of the race will accelerate progress and get them into the hands of consumers faster

• AI scientists are moving from vision to action in their research. During the 2010s, computer vision was the key problem in ML for a lot of top scientists and many spent their research specializing in it (e.g. Karpathy). Now many of the scientists that were previously working on computer vision have moved on to working on 'embodied AI' or creating AI that has the capacity to act in the world. An example would be Fei Fei Li, the professor (and Karpathy's tutor) who established ImageNet and is now working on WorldLabs.

• Nvidia is betting the company on humanoids. A lot of analysts think that Nvidia stock price is overvalued now because there are other specialized chips that are better than them in certain areas. For instance, there are various ASICs like the Groq chip that are much faster than H100s for LLM inference. But Jensen doesn't think that LLMs are the goal. He doesn't even think that building AI agents (requiring much more inference time compute) is the goal. He is dead set on general purpose robotics being the goal, as the slides from GTC 2025 show. You can use transformer-specific ASICs for chatbots, but training robots is far more compute-intensive since you need to be able to run the robots through years or decades of training in simulation. This is where Nvidia has the advantage - no one is better at networking together thousands of GPUs at the datacenter level for robot training.

Posts are 100% financial advice. You should go buy more Nvidia.