This week: In Seoul, Intel made a simple pitch to robot makers: stop wiring separate computers together and run a robot's reflexes and brain on one chip.

Elsewhere: Intel's open-source Physical AI Studio, and how to train your first robot policy without owning a robot.

The New Prime Signal

On October 14, Intel held its Robotics Workshop at the JW Marriott Hotel in Seoul under the theme "Physical AI in Action," a showcase for Korean customers, partners, and developers. The big idea, laid out by John Healy, who leads Intel's industrial and robotics business, and reported by ChosunBiz: run real-time motion control and AI on the same chip, the Core Ultra Series 3, instead of splitting them across separate systems. Intel's robotics page makes the same case in one line: "Real-time motion control and AI inference on a single platform." MoneyToday covered the factory-of-the-future aspect of the day.

What shipped: Less a new chip than a full kit built around one. Core Ultra Series 3 puts a CPU, a GPU, and an NPU (a processor tuned for AI) on a single chip. Per Healy, the CPU handles real-time control and system management while the built-in GPU runs the heavier AI models. On the software side, Intel showed Physical AI Studio, a tool for teaching robots by demonstration that covers data collection, model training, and evaluation, and OpenVINO Physical AI, which connects those models to cameras, sensors, and robot controls. Both sit inside Intel's Robotics AI Suite, whose latest release, 2026.2, landed on September 10. At the booths, Korean motion-control company Movensys showed AI and robot motion control running on a single PC with Core Ultra Series 3, Asus showed an industrial edge AI computer, and startup Circulus had a humanoid lifting and moving a box, running its own software on Intel Panther Lake chips inside hardware from Unitree.

Why it matters: A robot does two very different kinds of thinking. The AI part can take a beat to figure out what the box is. The motion part can't: a joint command that arrives late isn't a hiccup, it's a dent in something. Healy calls this "time determinism," finishing the work inside a set time, every time. Intel's argument is that putting both jobs on one chip means fewer boards and components, lower power use, and cheaper software development. There's strategy in it too. Asked how Intel compares with Nvidia's Isaac robotics platform, Healy picked "openness": train anywhere, fine-tune and simulate anywhere, then deploy on Intel. The caveats are real, though. This is Intel's pitch at Intel's event. Its robotics page claims 50% faster performance on a medium-sized vision-language-action model with a Core Ultra X7 358H than with Nvidia's Jetson Orin AGX 64GB, but that's Intel's own estimate, on Physical Intelligence's pi0.5 model. Its survey of 800 leaders in six countries found 60% expect to run robot fleets within five years, while only 40% have a formal strategy for humans and robots working side by side. Those are Intel's numbers too, and one chip doing everything still has to prove itself on real factory floors.

What's next: Healy said Intel came to Korea because opportunities with its customers and partners there are "growing rapidly," and that robots can keep gaining features through software after they're deployed. MoneyToday reports Intel is working with a KAIST spin-off led by professor Jang Young-jae on "dark factories," where AI runs the whole plant, and Jang said they plan to build real factories in the near future. The things to watch: named customers shipping production robots on the one-chip design, and benchmarks from someone other than Intel.

Robots need good timing more than they need genius. Intel's bet is that the company best known for PC chips can sell that sense of timing, plus open tools, to robot builders who would rather not be locked into one vendor. It's a sensible pitch aimed at a real problem. The proof will come from factory floors, not workshop slides.

The most hands-on piece of Intel's pitch is Physical AI Studio, and it's open source on GitHub, so you can poke at it tonight.

How it works: You record demonstrations of a task, usually by guiding a robot arm through it, then train a "policy," a model that learns to copy those moves. Physical AI Studio supports policies such as ACT, SmolVLA, and Pi0.5, plus the LeRobot policy zoo, and you can drive it from a Python API, a command line, or a web GUI. When the model is ready, you export it to OpenVINO, ONNX, or Torch, and Intel's physicalai runtime handles the cameras, the robot controls, and the inference loop on the robot itself.

Try it: The project's quick start installs with pip install physicalai-train and trains an ACT policy on LeRobot's simulated dataset aloha_sim_transfer_cube_human, so you can watch the whole pipeline run before you buy any hardware. If you get hooked, Intel's collect, train, deploy tutorial takes a real SO-101 arm from 50 to 100 demonstrations to a working pi0.5 policy. Fair warning: it asks for a GPU with at least 40 GB of VRAM for training, so start with the simulator.