Skip to content
Technology How it works Breeder — Hyperion Burner — Aegis Burner — MetroVolt AI-Native Architecture Magnets Fuel cycle Safety Roadmap
Solutions AI & Data Centers Defense & Government Grid & Baseload Neutron Detection Quantum
Learn Technical Library
Proof Publications Whitepapers Technical Library Open Science & Reproducibility The Honest Gates
Company About / Mission Leadership Environment Health & Safety Investors Careers Press Contact
3D Model
MetroVolt › Data-Center Power
Data-Center Power

Powering Training Clusters

A training cluster is a single large synchronous load held near full utilization for weeks; its firmness and ramp behavior shape how the burner must be integrated.

One computation, thousands of chips

A frontier training run couples tens of thousands of accelerators into one synchronous computation. They step together, exchange gradients over a high-bandwidth fabric, and stall together if any part stalls. Electrically this presents as a very large, very flat load with sharp collective transitions at the start and end of a run and at checkpoint boundaries.

Job startramp upSteady trainingflat, weeksCheckpointsbrief dipsJob endramp down

The flat, continuous middle is a good match for a firm base source — the burner supplies that plateau natively while it burns. The transitions are where power quality matters: a synchronized ramp of many accelerators is a fast, large step change, and the conditioning and battery layers must absorb it so the plasma is not asked to chase load transients it cannot follow quickly.

The stakes on continuity are high. Losing power mid-run can waste days of computation back to the last checkpoint, so the cluster cannot tolerate an uncovered burner trip. This is the 24/7 requirement in its sharpest form, and it is why a training campus leans on the full hybrid — burner base, battery bridge, grid backup — rather than on burner availability alone at 0.86-0.995.

Integrated correctly, the burner holds the training plateau while storage handles the transitions and outages, and checkpoints bound the worst-case loss if every layer somehow fails at once.

Content reviewed August 2026 · design-and-simulation stage