Heat recovery architecture

Put waste heat to work before rejecting it.

The thermal loop is central: recover heat from liquid-cooled computer equipment, transfer it into useful hot-water and drying loads, and only then route excess heat to the final rejection system.

Integrated water and heat-reuse system

The system diagram shows how cold water, wash water, filtration, hot-water storage, heat exchangers, liquid-cooled compute equipment, evaporative cooling and control valves integrate into one thermal and water loop.

Cooling details Environmental benefits

Supporting business-plan figures

Thermal energy reuse business-plan page — **Thermal energy reuse**Thermal-energy reuse, hot-water integration and heat-rejection design.

Hybrid solar installation business-plan page — **Hybrid solar installation**Business-plan planning image; the current PVT science reference array is 170 m².

Laundromat concept with rooftop solar, thermal storage and distributed compute

Heat-matched compute

Heat-sink capacity, not accelerator fashion, sets the useful local compute density.

In the current 120-host Los Angeles screening model, 6,160 B300-equivalent accelerators consume 63.49 GWh/year of IT electricity. Under the optimistic fully temperature-compatible heat-sink bound, 50.38 GWh/year becomes useful process heat: 79.4% of IT electricity and 93.4% of modeled host thermal demand.

The model also shows the penalty for oversizing beyond the heat sink. Increasing compute capacity by 50% lowers IT-to-useful-heat conversion to 56.7% and raises direct evaporative cooling water to 4,505 m³/year. Lower temperature compatibility has the same directional effect, so real deployment requires measured heat magnitude and temperature grade at each host. The architecture therefore distributes compute to thermal sinks and relies on dedicated dark fibre plus NVIDIA Spectrum-XGS to test whether those smaller sites can collectively replace the role of a mega data centre.

Open science downloads

Current papers, methods and model limits.

First page of the current heat-sink-constrained distributed AI infrastructure manuscript

Heat-sink-constrained siting of distributed AI infrastructure: energy, carbon and water impacts in Los Angeles

The current deterministic hourly study places 6,160 B300-equivalent accelerators across 120 representative car-wash and laundromat hosts and tests useful heat, carbon, water, climate and metropolitan-network constraints. Its network programme uses NVIDIA Spectrum-XGS as the scale-across layer over dedicated optical fibre and specifies fibre-spool and multi-site experiments to test whether distributed sites can match co-located clusters — the key experiment behind the project’s objective of proving that mega data centres can become obsolete as the default architecture for many workloads.

Current paper PDF Model + files

First page of the current mechanism-resolved fanless photovoltaic-thermal heat-rejection manuscript

Mechanism-resolved modeling and adaptive control of a fanless photovoltaic-thermal heat-rejection array

The current 170 m² roof-array study resolves passive and wetted mechanisms separately, quantifies the water-capacity trade-off, rejects array-wide ionic wind on energetic grounds and tests roof-area matching from micro facilities to hyperscale.

Current paper PDF DOCX Model + files

The goal: prove the mega data centre is no longer necessary.

Cloud Wash Laundry is being designed around a direct technical challenge to the conventional mega-campus model: place compute where useful heat already exists, then connect those heat-sink-sited AI factories with dedicated dark fibre and NVIDIA Spectrum-XGS scale-across networking so the distributed sites can operate as one metropolitan compute fabric.

The dark fibre is the dedicated optical transport; Spectrum-XGS is NVIDIA’s scale-across networking layer. The project’s proposed fibre-spool testbed and multi-site field pilot are intended to determine whether distributed sites can match co-located clusters on workload-level measures such as time-to-result and energy per token.

Measured next. Modeled now. Heat-sink constrained.

Current design-stage results combine the 120-host Los Angeles heat-sink-siting model with the updated fanless heat-rejection screening figure. They are screening outputs, not field-certified performance. In parallel, the network programme tests whether NVIDIA Spectrum-XGS scale-across networking over dedicated dark fibre can make geographically separated AI factories behave as one compute system. The mega-data-centre conclusion remains an experimental objective, not an already validated result.