Follow the heat

Follow the heat, from chip to town.

Every watt an AI server draws turns into heat. Air throws it away. Nexalus keeps it in water, hot enough to use.

No air path. The heat leaves as water at up to 60 °C.

Step 1 of 6 · ≈ 3 cm

The chip

Eight GPUs at full 600 W make 4.8 kW of heat, a few square centimetres at a time. Jets aim fresh coolant at the plate right above each hot spot.

≤ 80 °C Measured

Hottest GPU over one hour at full load (8 × 600 W)

1 +9.6% Cooler cards ran faster: 6,534 vs 5,961 tokens/s at 68 vs 87 °C (six cards, one box) Measured
Longer lifeLess capex
Source: Nova 8 one-hour soak, 5 Oct 2026, 175 back-to-back runs, hottest GPU 56 → 80 °C. Per-card figures: Aug 2026 development unit, stock clocks.
Where the server's heat ends up Sealed: 68% sold to a heat network, 29% to dry coolers, 3% to room air. Simulated
02 · Follow the heat