Nvidia makes 'physical AI' push
At GTC 2026 Nvidia framed AI as the new operating layer for robots, expanding Isaac tools for sim‑to‑real, announcing robotics foundation‑model work, and signaling a shift from chips to full AI‑optimized data‑centers — with Groq positioned as a complementary accelerator. Jensen Huang also highlighted China’s hardware strength, underscoring hardware supply‑chain dynamics that matter for robot builders. (theaiinsider.tech)
NVIDIA published new open robotics assets at GTC — naming NVIDIA Isaac GR00T and NVIDIA Cosmos as open models, releasing Isaac simulation framework updates plus Isaac Lab‑Arena and an OSMO edge‑to‑cloud workflow for robot training and evaluation. (investor.nvidia.com) The company announced explicit partner rollouts across industrial, surgical and humanoid robotics, naming ABB Robotics, FANUC, Figure, Agility, KUKA, Universal Robots, Yaskawa, CMR Surgical, Medtronic, Skild AI and Hexagon Robotics among collaborators using NVIDIA stacks. (investor.nvidia.com) On infrastructure, NVIDIA integrated a Groq‑derived LPU into new LPX racks (256 LPUs per rack) advertised with 128 GB of on‑chip SRAM and ~640 TB/s scale‑up bandwidth for low‑latency inference workloads. (datacenterdynamics.com) NVIDIA also expanded its Vera Rubin NVL72 platform into multi‑rack PODs aimed at “token output per watt” efficiency for continuous inference and agentic workloads, and paired that rack strategy with new Vera CPU and BlueField DPU offerings. (storagereview.com) The Groq arrangement that produced the LPU traces to a roughly $20 billion licensing/acquihire deal finalized in December 2025, a structure that has drawn regulatory scrutiny and prompted congressional questions even as NVIDIA integrates Groq IP and personnel. (cnbc.com) At GTC CEO Jensen Huang said export licenses and purchase orders exist for H200 shipments to China and that NVIDIA is “restarting” H200 manufacturing for those customers, a reversal covered by Reuters, Bloomberg and CNBC amid reports Beijing has greenlit some H200 buys. (msn.com) NVIDIA tied these robotics and datacenter moves to scale deployments — announcing a DRIVE/Alpamayo robotaxi partnership with Uber that will start in Los Angeles and San Francisco in 2027 and expand to 28 cities globally by 2028 under the company’s DRIVE Hyperion stack. (investor.uber.com)
