Drop-in · Plug-in · Seal

Watch the plant disappear.

The same 1 MW of AI in Frankfurt, modelled four ways. One server for every step; the Plug-in kit and sealed rack are in design.

The PUE paradox

At the drop-in step, PUE stays at 1.42 while energy falls and tokens per kWh rise 22%. PUE cannot see the gain, so we report energy and tokens per kWh.

PUE±0%
Energy−3%
Tokens/kWh+22%
Simulated
Model assumes Nova 8 at ~31 M tokens/kWh, +22% on the air reference (estimate; server power not yet wall-metered). Both are whole-server figures, so each server’s fans are counted once. Global average PUE is 1.54, flat for six years (Uptime Institute 2025, reported). ISO/IEC 30134-2: PUE "should not be used to compare different data centres."
More tokens, less opexZero unplanned downtime

03 · Run · Seal the rack

1.11
PUE −22%
Simulated
27.9M
Tokens per grid kWh +56%
Simulated
7,878MWh/yr
Facility energy −27%
Simulated
17
Racks for 1 MW −76%
Simulated
Where every kW goesFacility draw 1,124 kW · 4,806 MWh of heat reused a year
Model: Nexalus greenest-data-centre model, 1 MW IT in Frankfurt, 80% load. Tokens per grid kWh = whole-server tokens per kWh (25.4 M air, 31.0 M Nova 8, estimate) × IT power ÷ facility power, so fans count once. 70% of captured heat sold assumes a district network large enough to take the heat all year; a single seasonal building takes 20–35%. Homes = yearly heat use at 12 MWh a home. Evaporative towers at 1.6 L per kWh of heat rejected. Overheads are typical fractions per architecture, scaled by climate; illustrative, not a site survey.
02 · Watch the plant disappear