The bottleneck in physical AI is infrastructure

Britain has world-class talent, cutting-edge research and a growing set of startups building machines that can work in warehouses, factories and on the road. However, many of those teams lack the infrastructure to turn a working prototype into a fleet. Infrastructure increasingly sets the pace of robot model development, because it has to store, move and process the vast amounts of multimodal data these machines generate.

What physical AI is, and why it is different 

Physical AI describes autonomous systems that perceive, reason and act in the real world. It is applicable across many industries, including healthcare, transportation, manufacturing and logistics. In healthcare, robots can help automate experiments in the laboratory, and in transportation, autonomous vehicles can improve safety and accessibility.  

Unlike a generative AI chatbot, a robot or autonomous machine takes action that changes its environment, and that change feeds the next decision. The data is different too: video, LiDAR and 3D geometry, rather than text. Finally, the stakes are higher than for a chatbot hallucination. A poor control decision on a robot can damage hardware or put people at risk. 

Why the loop needs purpose-built infrastructure 

In physical AI, iteration speed is the competitive advantage. The goal is not simply to train a bigger model but to shorten the time and effort needed to go from observation to a validated improvement. Every stage of this “learning loop” depends on infrastructure: 

• Compute: Training, model experimentation and parallel simulation need large GPU clusters, fast interconnects and reliability. A single node failure can stall a long training run, so automated recovery matters.  

• Storage: Multimodal robot data overwhelms conventional cloud storage. Keeping GPUs utilised takes a tiered approach that includes reliable parallel filesystems, highly scalable object storage and high-performance local storage. 

• Cloud services: We estimate that teams routinely spend 30-40% of their time on compute integration work rather than improving robot behaviour. Orchestration that links simulation, training and evaluation helps engineers achieve results faster. 

Rhoda, a Silicon Valley robotics innovator, narrowed the gap between prototype and production robots by doubling its compute capacity and using petabytes of data storage adjacent to that compute. NormaCore, based in Europe, can train vision-language-action (VLA) models and deploy updated policies without stitching together multiple infrastructure vendors. https://nebius.com/customer-stories/roboforce 

Where teams stall, and what the UK needs to get right 

Physical AI teams such as Rhoda and NormaCore need reliable training compute, a way to generate and curate synthetic data, and simulation that runs alongside training. Building that in-house means capital expenditure before a single model is trained, and engineering attention that would be better spent on the robot.  

Hubs across the UK that are advancing robotics in surgery and farming will drive innovation forward. To build on these efforts and lead in physical AI, three things matter even more. First, access to current-generation accelerated compute on flexible terms, so capacity is available to scale with demand. Second, integrated tooling for simulation, synthetic data, training and evaluation, so teams spend their time on robot behaviour instead of compute plumbing. Third, openness: stacks built on open frameworks and shared improvements help the whole physical AI ecosystem move faster. 

The Nebius Physical AI Living Lab 

This is the thinking behind the Physical AI Living Lab, a programme run by Nebius and NVIDIA for robotics teams in the UK and Europe. Robotics startups get access to NVIDIA’s physical AI stack, including NVIDIA Cosmos, NVIDIA Isaac Sim and NVIDIA Isaac GR00T, wired together and running on Nebius AI Cloud with up to 16 NVIDIA RTX PRO 6000 Blackwell GPUs. Applications run through the NVIDIA Inception pipeline. The first cohort is underway.  

At the centre of the programme is the Nebius Physical AI Workbench. It turns NVIDIA and third-party physical AI tools into composable building blocks that engineers and agents can wire together into workflows. In these fully automated workflows, engineers can use natural language to interact with robots, in the same way people use generative AI chatbots for tasks in the digital world.  

Early signals from the first cohort are positive. One company has validated multiple physical AI workflows with GPU utilisation above 90%, helping it iterate faster. Another is running reinforcement learning across thousands of simulated robots on Nebius compute, saving time and improving safety before robots are deployed in the real world. A third generates simulated environments directly from its own collected data, enabling testing of robots on edge cases that are difficult to produce otherwise. Participants are also contributing back: a fix from one team has already been merged into the open-source Workbench repository. These results suggest that access to accelerated compute, integrated tooling and an open ecosystem can drive faster adoption of robots that are useful and safe in the real world.  

Closing the gap 

Physical AI will move from research to everyday deployment at the speed of its learning loop, and that loop is only as fast as its slowest piece of infrastructure. If the UK wants its robotics innovators to scale, it needs to make fast, integrated, reliable infrastructure the default. Robotics startups in the UK and Europe can register their interest in future cohorts at nebius.com/physical-ai-living-lab. Further information on Nebius for physical AI is available here.   

About the authors 

Russell Lamb is the UK Government Affairs Lead for Nebius, covering their work on policy and engagement across the UK’s tech ecosystem. He has worked in technology policy and public affairs for a number of years, as well as spending time in the civil service designing tech policy in different departments.  

Akshai Parthasarathy is Director of Solutions Marketing for Physical AI at Nebius. His areas of focus include robotics innovators and physical AI software providers. He was previously at Oracle Cloud Infrastructure (OCI), leading Product Marketing for AI infrastructure. He earned an MBA from UC Berkeley and engineering degrees from Georgia Tech.  

About Nebius 

Nebius, the AI cloud company, is building the full-stack platform for developers and companies to take charge of their AI future; from data and model training to production deployment. Founded on deep in-house technological expertise and operating at scale with a rapidly expanding global footprint, Nebius serves startups and enterprises building AI products, agents, and services worldwide. 

For more information please visit nebius.com.