forty percent of the power just keeps it cool

Problem statement

U.S. data centers consume approximately 2.5% of national electricity, and cooling systems account for 30–40% of that energy — a proportion that is growing as AI workloads drive chip power densities beyond what conventional air cooling can handle. Current GPU accelerators (NVIDIA H100, B200) dissipate 700–1,000W each in dense configurations, creating heat fluxes that overwhelm traditional computer room air conditioning (CRAC) systems. The industry is approaching a thermal wall: cooling is becoming the binding constraint on data center capacity, not compute hardware or electrical supply.

Why this matters

Global data center energy consumption is projected to more than double by 2030, driven by AI training and inference workloads. If cooling energy remains at 30–40% of total facility power, the cooling burden alone could exceed the total current energy consumption of many small nations. The Power Usage Effectiveness (PUE) metric — total facility energy divided by IT energy — has plateaued at ~1.3 for the best operators, meaning 23% of energy goes to overhead (mostly cooling). ARPA-E's COOLERCHIPS program targets total cooling energy below 5% of IT load at any U.S. location, which would represent a 6–8× reduction from current practice.

What’s been tried and why it hasn’t worked

Air cooling (the dominant approach) is reaching its physical limits as chip power densities exceed 100 W/cm². Rear-door heat exchangers and hot/cold aisle containment improve air cooling efficiency but don't change its fundamental heat-transfer limitations. Single-phase liquid cooling (cold plates with water or coolant loops) can handle higher heat loads but requires plumbing infrastructure, leak-proofing, and doesn't fully eliminate fans. Two-phase immersion cooling (submerging servers in dielectric fluid that boils on hot surfaces) offers excellent heat transfer but requires redesigned server hardware, specialized fluids with uncertain environmental profiles (some are fluorinated compounds), and creates maintenance challenges — technicians can't quickly access components in a bath of fluid. Direct-to-chip (DtC) microfluidic cooling has shown promise in labs but faces manufacturability and reliability challenges at scale.

What would unlock progress

COOLERCHIPS ($42M, 15 projects) funds transformational cooling approaches targeting the 5% energy threshold. Promising directions include: chip-integrated microfluidic cooling channels fabricated directly into semiconductor packages (eliminating thermal interface materials), two-phase cooling with non-fluorinated working fluids, waste heat recovery systems that capture cooling energy for heating or power generation (turning the cooling burden into an asset), and AI-optimized cooling control systems that predict thermal loads and pre-position cooling capacity. ARPA-E expects 90% of funded approaches to fail but the 10% that succeed to be industry-transforming.

Entry points for student teams

A team could design and test a small-scale two-phase cooling loop using a non-fluorinated working fluid for a representative chip thermal load, measuring heat transfer coefficients and comparing against published immersion cooling performance. Alternatively, a team could develop a predictive cooling controller using ML models trained on publicly available data center thermal data. Thermal engineering, mechanical engineering, and computer science skills apply.