1U AI Server -- dual-GPU inference box. Distinct from a hosted-game server in that its whole economy is a compute-rental curve rather than a player-count simulation.
2x GPU, 24GB VRAM each, 100 TFLOPS each (200 total). 400W typical draw, 600W max -- this dwarfs everything else in the roster short of a fully-loaded crypto rig, so budget PDU capacity accordingly before racking one.
Real compute-rental economy: set a rental price, and utilization and revenue respond to it the same way a hosted game's player count responds to conditions -- price it too high and utilization drops; too low and you're leaving money on the table even at full utilization.
Because typical draw is so close to max draw (400W vs 600W, versus most other devices where typical is a small fraction of max), an AI server under real load runs consistently hot rather than spiking occasionally -- factor that into degradation/thermal planning more than you would for a server or router.
Tip: this is one of the few devices where the *typical* wattage figure, not the max, is what you should size PDU/UPS capacity against -- it's supposed to run near that number whenever it's actually earning.
See also: the Active Services Monitor for live utilization and revenue at a glance.