ThermoLoop: AI Data Center Thermal Recycling

Problem

AI data centers consume 1 to 2% of global electricity and reject large quantities of heat as waste. Conventional cooling systems treat that thermal output as a liability. A 20 MW IT load produces roughly 20 MW of thermal output, most of which is vented to the atmosphere unused. The iMasons challenge asked teams to propose an architecture that changes that.

Approach

ThermoLoop is a closed loop thermal recycling architecture that captures waste heat at every stage and routes it toward useful work. The system operates across four stages.

• Stage 1 (Direct to chip cooling): Two phase cold plates mounted on server CPUs and GPUs. Coolant exits at 50 to 65 degrees Celsius. Up to 82% less cooling energy compared to air cooling.
• Stage 2A (Adsorption chiller): Silica gel and water adsorption cycle produces chilled water at approximately 7 degrees Celsius. Reduces mechanical cooling load by approximately 80.5%. COP 0.3 to 0.5.
• Stage 2B (Organic Rankine Cycle): ORC turbine converts remaining high-grade waste heat to electricity at 8 to 15% thermal efficiency.
• Stage 3 (AI thermal dispatch): Intelligent control layer routes thermal output in real time to four destinations: adsorption chiller, ORC generator, thermal storage buffer, or community heat export loop, based on demand signals.
• Stage 4 (Community heat export): District heating, greenhouse agriculture, and aquaculture. A 20 MW IT load produces roughly 20 MW thermal output, enough to heat approximately 10,000 to 15,000 homes.

Results

My Role

Co-designed with Sam Wolpert. I contributed to the four-stage system architecture, stage integration logic, AI dispatch framework, and the pitch presentation delivered to iMasons judges. Both team members shared analysis and writing equally.

Reference Material

Skills Demonstrated