Compute Is Destiny
Robot brains will be trained on infrastructure being built right now. At Raise Week 2026 we chased the physical AI compute question – and found the builders disagreeing, honestly.
What does physical AI compute actually need – more GPUs, better architectures, cheaper energy, or smarter data? The frontier-lab panel at Raise Week 2026 could not agree, and that disagreement was the most honest moment of the week. Part three of our Raise Week field notes.
The Physical AI Thread · Field AI at the Frontier-Lab Panel
The training-at-scale panel – Shubho Sengupta (Axiom Math), Frank Hutter (Prior Labs), Eiso Kant (Poolside) and Ali Agha (Field AI), moderated by Felicis' Feyza Haskaraman – could not agree on the bottleneck: compute, architecture, energy or synthetic data. That disagreement was the most honest moment of the week.
For robotics, Ali Agha was the essential voice. Field AI builds foundation models for robots operating in complex, safety-critical environments like construction sites, and his argument cut against the scaling orthodoxy: robotics does not have internet-scale data and never will. All the teleoperation in the world is a drop in the ocean, and generalisation in the physical world is multi-dimensional – not the one-dimensional, word-based kind LLMs solved. Simulation data, video data, teleop data – Field AI uses everything it can get – but data alone will not produce general-purpose robots. His unlock: a next generation of transformers with built-in physics, architectures that need less data because they already know how the world behaves.
"Architecture is a moat around your data. You land-grab the market, you create a data flywheel – the architecture gives you the head start."
Eiso Kant of Poolside grounded the other end: model capability is limited by compute, but the real bottleneck turned out to be data centres and energy – 200 MW costs about $3 billion before power – so Poolside started building its own. And inference has to get faster: robots and agents need models that stream 10,000 tokens per second, not 30–40.
Frank Hutter's Prior Labs works on tabular foundation models – and there are only about 50,000 tabular datasets available online. Not every corner of AI has an internet to scrape. Robotics knows the feeling.
New! 2026 Humanoid
Robot Market Report
198 pages of exclusive insight from global robotics experts — uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
now Google DeepMind
Compute Is Destiny · Where Physical AI Compute Gets Built
The infrastructure sessions – Cloudflare, Vast Data, Credo, Clockwork, Solidigm, moderated by SemiAnalysis' Jordan Nanos – shared one refrain: everyone is sold out. Demand for compute looks endless; the hesitation is only about whether the exponential can continue. There is no lack of capital – the question is what margin you can charge, and whether the end users ("offtake") keep subscribing.
Two details stood out. Cloudflare says half the traffic on its network is now agentic, and it has gone all-in on isolates instead of containers – think flexible workspace instead of office space – to squeeze ten agents per worker out of ten billion CPUs. And the neoclouds are learning the hard way: 100,000 GPUs mean 100,000 km of fibre with no time to test it, GPUs rented at $6 an hour on leverage, and service rates on GPUs as high as 20% – a thousand GPUs can lose one to four hours a day to unutilised capacity and infrastructure failures. Reliability, as Credo put it, is the north star.
The robotics read: this is the layer robot foundation models will be trained on. When the people selling the shovels say demand is endless and one GPU in five needs service, a good part of the price and pace of robot-brain training is being set right here.
We went to Raise Week to listen for robots – and kept hearing the stack being built beneath them.
This is part three of our Raise Week 2026 field notes. Start with The Nordic Night and The AI Agent Shift – or read our Machina 2026 report from Station F.
New! 2026 Humanoid
Robot Market Report
198 pages of exclusive insight from global robotics experts — uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
now Google DeepMind
