Physical AI will be this year's Christmas present

What was discussed

We took the whole of Woomul, a cultural space in Hapjeong with an outdoor yard, two indoor floors and a rooftop. Bloom usually meets in something closer to a lecture room and had run an indoor event the day before, but this was the opening of a hackathon and called for somewhere else. Even on a Friday evening people arrived early for the 5pm check-in. We ate, made introductions, and started at half past five. About a quarter of the room was here for the first time.

The hackathon is one Nebius runs with NVIDIA across 20 cities. It opened in Tokyo two days earlier and runs through Da Nang and Seoul on to London, New York and San Francisco. Everything happens online, with submissions closing at 2am Korean time on 31 October, so this was a starting line for forming teams and getting hands on the tools rather than a day for producing anything. The invitation said as much: no demos, no pitching.

There are two conditions for entering. What you build has to run on Nebius cloud, and it has to use at least one NVIDIA open-source model. Four tracks: coding and agentic engineering, apps and agents, personal AI, and physical AI. Prizes come to more than $50,000 in total. The physical AI track carries one more requirement, which is at least a minute of a real robot moving in the submission video. Simulation footage will not do. Most of what was said on stage that evening ran into that condition.

The first session was Guan-ru Huang, a cloud solutions architect at Nebius. Asked who had heard of Nebius, not a hand went up, which he said was all the better. What he emphasised was the structure rather than the share price. Hyperscalers built their data centres for general workloads rather than AI; GPU rental companies at the other end hand over GPUs with no software stack. Nebius, he said, builds data centres for AI from the ground up and puts its own cloud software on top.

The first thing customers look at when sourcing GPUs is how quickly they can get them. ML scientists panic, he said, when they have been hired and there is no GPU to train on. Training runs now last weeks or months, so a cluster dropping mid-run means starting over, which puts stability next. And ten GPUs today becomes thousands as a business grows, so whether you can grow inside the same cluster matters too.

Then Token Factory, their inference platform. Developers today either use a closed API or self-host. The closed API is easiest to start with and eats the margin as a service grows; self-hosting needs a team dedicated to running the inference cluster. What he identified as the shared limitation was something else: neither gives you a way to keep improving the model with your own service's traffic. Token Factory builds that loop into the product, starting inference on an open-source model, exporting the traffic, showing what customers actually ask, then retraining and deploying to the GPU type and region you choose.

Tavily, which Nebius acquired, closed the session. Ask a model the current score and it has to go to the web, because nothing anywhere has been trained on what happened two seconds ago. Search results come back in a shape agents struggle to read, so Tavily sits in between: it returns the relevant URLs, pulls the content out of them, and follows the relationships between them. Running through accuracy, relevance and speed, he said, is freshness. Ask about the latest upgrade and get documentation from three years ago and you have nothing.

The second session was Dhruv Diddi, who leads the physical AI ecosystem. His slide setting digital and physical AI side by side made the point cleanly. Digital AI outputs tokens on a screen, so a hallucination is just wrong text and an error costs nothing. Physical AI outputs motor torque, so a hallucination becomes a collision and an error is catastrophic. One instruction not to drop the glass has to end up as joint angles and torque.

For why robots are not yet everywhere, he pointed to a three-body problem. The factory that collects data and trains, the matrix that runs simulation, and the edge the model ships to are three different computers: a DGX-class cluster, an RTX-class server generating synthetic data, and an onboard computer that has to react in milliseconds. Data collection sometimes does not match simulation, and simulation sometimes does not match deployment. The impressive videos online, he said, are the times everything did line up.

Training data comes from three places. Video scales best but carries no force or joint information. Simulation produces millions of episodes overnight on approximate physics. Teleoperation pairs action and state perfectly but yields five to fifty episodes an hour. At that point he produced an open-source teleoperation arm and sent it around the room: $200 online, 3D printed so you can repair it yourself, and the most scalable robot available right now. The dancing robots online are all reinforcement learning, since nothing is being handled. Getting one to hold a cup of water takes imitation.

The panel added Andy Lee of NVIDIA and Yubin Park, CBO of Dyden Robotics. Dyden builds quadrupeds that walk the walls and ceilings of steel structures on magnetic feet to weld. They have passed validation with Samsung Heavy Industries and entered delivery, and are developing an internal block welding robot with HD Hyundai. Asked what breaks first when thorough simulation meets reality, Park was specific. You build the simulator with flat floors and clean angles, but in an actual shipyard welding scatters metal chips across the floor, and a robot that steps on one slips. Without the real world, he said, you cannot build any solution to it.

Asked whether the AI or the hardware is harder in physical AI, Andy Lee said it is half and half. Models improve fast, but the moment you put one into the world the lighting is different, the environment is different, and variables you never accounted for pour in. Hardware has to catch up, and that is the stage we are at. For why Korea matters strategically he used a number: 1,220 industrial robots per 10,000 manufacturing workers, first in the world and more than nine times the global average. Asked what he would do differently founding again, Park did not hesitate. Do the hardware. Optimise only the software layer and it is far too easy for a customer to strip it out and swap in another vendor's.