GPUs pull their clocks back when they run hot. Air cooling struggles as the room warms and the rack fills; a liquid loop holds every GPU at a steady temperature, so the gap widens exactly when conditions get tough.
Illustrative model, not a measurement. Anchored to measured Nova 8 GPU temperatures (71–80 °C at full load) and the published air-cooled result (5,691 tok/s per GPU); throttling assumed above 87 °C.
Pick a model, precision, context and users. See the memory it needs and the server that fits.
Open →TOOL 02Chat with a model on Nova 8 and watch tokens per second in real time.
Open →TOOL 03Take Nova 8 apart in 3D. Every component, labelled and explained.
Open →TOOL 04Fill a rack with Nova servers and see power, heat, tokens and the cooling you need.
Open →TOOL 05Own vs rent: payback, cost per million tokens, and a report for procurement.
Open →TOOL 07How many homes your AI could heat, and what that heat is worth.
Open →TOOL 08Place a life-size Nova 8 in your own room with your phone.
Open →