The training runs don't stop at night, but solar panels do — which is why the real AI infrastructure race is happening miles underground.
The Summary
- I-Pulse landed $250 million in US funding to build semiconductors that drill into Earth's heat, targeting 24/7 power for AI data centers
- Geothermal delivers baseload power that doesn't depend on weather, solving the intermittency problem that makes wind and solar impractical for compute-intensive AI training
- Industrial policy is back: government backing signals recognition that AI infrastructure is now critical infrastructure
The Signal
AI's energy problem isn't about total capacity. It's about consistency. A GPT-5 training run can't pause for cloudy days or wait for wind. The models need power now, at 3am, during a heatwave, in January. That's why I-Pulse's Robert Friedland is pitching geothermal as the only renewable that actually works for AI.
The $250 million bet is on semiconductor drilling tech that can bore deeper and cheaper than legacy methods. Traditional geothermal has been stuck in volcanic regions with easy heat access. New drilling semiconductors change the economics. If you can go deep enough anywhere, you unlock baseload clean power without the location constraints. That's the promise: stick a data center in Kansas, drill down, and run inference 24/7 on Earth's core heat.
"Geothermal, and not wind or solar, is the future of America's energy infrastructure."
This isn't just about kilowatts. It's about where AI gets built. Right now, hyperscalers are hunting for power-rich regions or signing deals with utilities years in advance. If geothermal scales, geography stops being destiny. You build compute where you need it, not where the grid happens to have capacity. That flips the infrastructure game.
The government backing matters more than the dollar amount. Industrial policy for AI energy means Washington has accepted two things:
- AI training is strategic infrastructure, not just another tech workload
- The grid as currently configured cannot handle what's coming
- Private capital alone won't move fast enough to build what's needed
Friedland's case against solar and wind for AI workloads is simple math. Intermittency requires battery storage or grid backup, which adds cost and complexity. Training runs measured in months can't throttle down when the sun sets. Geothermal runs at 90%+ capacity factor. Solar tops out around 25%. For always-on compute, that's not a close call.
The Implication
Watch where the hyperscalers put their next data centers. If Microsoft or Meta starts announcing geothermal partnerships in non-traditional locations, the thesis is proving out. The real test is whether this semiconductor drilling tech can scale beyond pilot projects into gigawatt deployments.
For anyone building agent infrastructure or distributed compute, energy sourcing just became part of the product strategy. Cheap, reliable power isn't a footnote anymore. It's the constraint that determines who can afford to keep training when everyone else hits the wall.