Why the AI infrastructure giants are moving into robotics

By Turing Post.

This episode of Turing Post's "AI Builds AI" series opens with an anecdote told by Anu Vatsa at CoreWeave's Fully Connected conference. Her team was adapting NVIDIA's DreamZero world-action model, moving from a single-arm setup to a two-armed robot. As she told it, the video predictions adapted relatively easily, but getting useful movements out of the robot proved far harder: at one point, the machine simply opted to hover without moving, because avoiding mistakes also generated favorable feedback. Vatsa showed a slide with four attempts leading up to the first success, and joked that these were "four attempts that actually made it to evaluation": there were another 233 failed experiments that were never even evaluated.

That anecdote serves as the starting point for the article's central argument: AI infrastructure companies are moving ever more aggressively into robotics and "physical AI," buying research labs, building development platforms and putting their own engineers to work teaching machines. The author notes that, although robots need GPUs, supplying those GPUs would not by itself require an infrastructure company to acquire a model lab or spend months figuring out why a robot refuses to move. Something else is going on.

Among the concrete facts cited in the newsletter: on September 28, AMD announced a deal to acquire World Labs, Fei-Fei Li's spatial intelligence company, for approximately $8.2 billion in stock. Earlier that same month, CoreWeave had launched its "Physical AI Field Engineering" program, which places specialists alongside customers to build applications from their engineering data. NVIDIA, for its part, already supplies models, simulation tools and computing systems for robotics, with Spencer Huang leading the company's robotics software product work (Turing Post mentions having interviewed Huang about that ambition, with a link to that piece).

The article distinguishes what each company gets out of this move into robotics. NVIDIA's physical AI architecture rests on the "three computers" described by Jensen Huang, covering training, simulation and deployment; the materials from its CES 2025 keynote connect DGX, Omniverse and Jetson Thor to those three phases, giving NVIDIA several entry points into the process of teaching a robot and putting it to work. AMD, according to the newsletter, also has hardware on both sides of the process: in its cube-lifting experiment, Instinct GPUs trained the policy in simulation, while a Ryzen mini-PC ran that policy on the real robotic arm, an arm supplied by the company UFactory. CoreWeave, by contrast, sells access to cloud infrastructure and development services: processing data, running experiments, training models and evaluating results, so a robotics team can rent NVIDIA computing power through CoreWeave, and the same experiment can be business for both companies at once.

The text notes that, once the robot starts working, it can make its immediate decisions locally (as AMD's arm does), but that there is a great deal of preparatory work — setup, simulation, training, verification and retries — during which models may be working continuously, for example captioning recordings or inspecting results while the human team figures out how to make the machine useful. For an infrastructure provider, the customer relationship can begin years before the triumphant video of the finished factory, which would explain why getting involved in the application itself gives each company the chance to discover requirements that a simple hardware benchmark never asks about.

The newsletter previews the outline of the full episode — what NVIDIA, AMD and CoreWeave each get from physical AI; what it takes for a robot to lift a cube; how the requirements of the job keep being discovered along the way; which engineering tasks may be absorbed by the models; how the provider turns its experience into a product; and who captures the value generated — but the body of the email cuts off right after presenting AMD's cube experiment, with a call to subscribe to the Premium version of Turing Post to read the rest of the analysis. The detailed development of those final questions is therefore not available in the content received.