Accelerator Cluster Power Profile

An accelerator cluster draws a high idle floor plus fast, correlated swings during computation; understanding that profile sets the conditioning and storage the burner needs.

Floor plus fast swings

Modern accelerators do not sit at a single power level. They idle at a substantial floor for memory and interconnect, then surge during dense compute phases and relax between them. When thousands run one synchronized job, these swings correlate, so the campus sees large, fast steps rather than smooth averages.

Two properties follow. First, the floor is high and continuous, which is exactly the flat base a burner supplies well. Second, the swings are fast and synchronized, faster than a fusion plasma should be asked to follow. So the architecture assigns the base to the burner and the transient response to power conditioning and the battery, which can source and sink current in milliseconds.

The correlated nature of the swings also raises the ride-through requirement. A whole cluster stepping together is a large disturbance, and the ride-through layer must be sized for the aggregate, not for one rack. This is a normal data-center engineering task, but it is amplified by AI clusters' unusually synchronized behavior.

None of this changes the reliability picture: the cluster still cannot tolerate an uncovered loss of the base source, so the burner's availability gate is handled the same way — by the hybrid — regardless of the fine structure of the load.

• High continuous idle floor suits a firm base
• Fast, correlated compute swings exceed safe plasma ramp rates
• Conditioning + battery handle transients, burner handles the floor
• Ride-through must be sized for aggregate, synchronized swings