The hydrogen and liquid cooling shift in AI infrastructure

Hydrogen Fuel Cells and Power Reliability

I see the shift. Hydrogen fuel cells generate electricity through a chemical reaction between hydrogen and oxygen. This process produces only water vapor. Microsoft, Caterpillar, and Ballard Power Systems demonstrated a 1.5 MW hydrogen fuel cell system in Wyoming to prove that large-scale fuel cells can meet the 99.999% uptime requirements necessary for hyperscale data center operations. Proton Exchange Membrane (PEMFC) fuel cells provide the fast startup needed for backup power. Solid Oxide Fuel Cells (SOFC) provide the 60% electrical efficiency required for primary power.

In a combined heat and power (CHP) setup, fuel cell waste heat drives absorption chillers. This process increases total system efficiency from 50% to 80%. For a 100 MW facility, they use absorption chillers to reduce PUE from 1.44 to 1.30 and free 7.7 MW for additional computing load. A 2.5 MW FuelCell Energy system consolidates exhaust into a single outlet to reduce installation costs. These systems provide up to 450 tons of cooling per 2.5 MW of electric output. Operators apply thermal output from a 58.8 MW fuel cell installation in Korea to data center needs to generate more than 10,000 tons of chilled water. Hydrogen production through electrolysis using renewable energy provides a path to near-zero lifecycle emissions. The hydrogen fuel cell market for data centers reached $0.74 billion in 2025. This market reaches $4.64 billion by 2036, growing at a CAGR of 16.8%. The hydrogen-powered data center market reached $3.2 billion in 2025 and reaches $18.9 billion by 2035. This market grows at a CAGR of 19.4%.

Thermal Management and Liquid Cooling

Liquid cooling works. Air cooling reaches its physical limit at 41.3 kW per rack. I find the $2-3 million per megawatt cost for liquid cooling retrofits hard to ignore. You already know that AI training requires more power than traditional cloud services.

Liquid cooling manages the massive heat from NVIDIA H100 and AMD MI300X chips. These chips generate over 700W per GPU. Air cooling fails at these densities. Direct-to-chip cooling removes 70-80% of heat using cold plates. This method allows operators to increase ambient temperatures to save power. Immersion cooling submerges hardware in dielectric fluid. Two-phase immersion uses fluids like 3M’s Novec that boil at 34-56 degrees Celsius.

Air carries 3,300 times less heat per unit volume than water. A single gallon of water absorbs the same heat as 3,000 cubic feet of air. Operators move 1 kW of heat with air using 100 cubic feet of airflow with a 10 degree temperature rise. This creates 95 decibels of noise. Liquid cooling provides 60x better heat transfer than air. Liquid cooling market reached $5.52 billion in 2025. This market reaches $15.75 billion by 2030. Liquid cooling implementation covers 22% of data centers as of 2025. Immersion cooling can lower energy expenditures by 20%.

Market Scaling and Strategic Deals

Hydrogen scales fast. Bloom Energy signed a deal with Oracle for up to 2.85 GW of power. Utility AEP also signed a 1 GW supply agreement with Bloom Energy. Equinix uses over 100 MW of fuel cell capacity across 19 US sites. The $5 billion partnership between Brookfield Asset Management and Bloom Energy funds fuel cell projects for AI data centers.

The transition from backup to primary power accelerates as companies bypass grid constraints. PEMFC technology remains the choice for diesel generator replacement. SOFC technology leads in primary power applications. AI-related data center infrastructure needs $5.2 trillion by 2030. AI-optimized servers use 44% of total data center power by 2030. AI-optimized servers grow 30% each year. The demand for AI-related data center capacity reaches 156 GW by 2030. US data centers used 176 TWh in 2023, and this share reaches 7.8% by 2025. Will the industry resolve the high capital expenditure for liquid cooling retrofits before the next wave of AI chips arrives?