Nvidia wants to own every robot brain on earth
Jensen Huang’s silicon empire already decides who trains the biggest ai models. Now he wants to crawl inside the skull of every humanoid that will ever walk a factory floor.
The gpu king looks for a second crown
While rivals chase chatbots, Nvidia is quietly wiring European chip giants Infineon, NXP and STMicroelectronics into a supply chain purpose-built for legged machines. The goal: supply the cortex, let partners handle the clumsy limbs. Call it the “Detroit model” for robots—Nvidia ships the engine, someone else bolts on the chassis.
Inside the Santa Clara headquarters the project carries a codename that sounds like a Bond villain: NemoClaw. Engineers describe it as a hardened fork of OpenClaw, the open-source robot stack circulating in Chinese labs. The twist is trust: every motor command must pass through a locked-down driver signed by Nvidia’s cloud. No signature, no torque. Robotics departments hate it, procurement departments love it—because when a $200 k humanoid collapses, legal wants a single throat to choke.
The platform is already running inside Figure 02, the biped that BMW parades around its Spartanburg plant. Cameras feed 30 GB/s of depth data into a Jetson Thor board; inference latency is 12 ms, fast enough to stop a 90-pound arm before it punches a windshield. BMW won’t let reporters film the failures, but line workers gossip: the robot still drops a door panel every 47 minutes. Nvidia calls that “statistically acceptable” for a beta.
Why now? because the parking lot is full of bodies
Autonomous-car dreams stalled at Level 3, leaving Nvidia with warehouses of Orin chips that can fuse 360° vision but have nowhere to go. Repurposing them for bipeds is cheaper than designing from scratch. The math is brutal: a single robot needs one-fifth the silicon of a robo-taxi, yet sells for triple the price because factories will pay ransom wages to avoid 3 a.m. shifts.
Still, the skepticism inside the robotics guild is volcanic. “Walking is solved; folding a fitted shirt is not,” mutters a Toyota engineer who asked not to be named while describing his company’s internal tests. “Nvidia gives us a Ferrari brain to steer a wheelbarrow body.” Production cost for a humanoid hovers around $150 k; leasing starts at $8,300 a month—more than a veteran welder’s salary in Guangzhou.
Huang counters with volume. At the last GTC he flashed a slide: “1 million robots by 2029” stamped over a black-and-yellow GPU shroud. The audience cheered, but Wall Street analysts scribbled the same note: “Only if someone figures out use cases beyond viral demos.”
For now, Nvidia’s revenue line for robotics is a rounding error—$296 million last quarter, 1.4 % of total sales. Yet the R&D budget for “physical ai” tripled year-over-year. The bet is classic Jensen: get in early, own the SDK, then tax every servo that twitches. If the robots stumble, Nvidia still sells the supercomputers that train their replacements. Either way, the house wins.
The company that taught servers to hallucinate now wants to teach metal to walk. Factory bosses will decide whether that prophecy ships on time or ends up as scrap next to the graveyard of 3-D TVs. One thing is certain: whoever owns the brain writes the warranty.