Data Centers in America: Digital infrastructure, energy reality, and the rise of closed loop systems, Greg Barnett, Oil and Gas 360 (2026).
Uses tensor processing units, rather than GPUs. TPUs have slightly lower energy usage.
Since 2015, google has been making custom TPUs for things like powering Google Search
Relatively little. CPUs are mainly used. Single racks use about 10-15 kW
Because of the lower power usage, they can often get away with air cooling. Alternatively, there's mild evaporation-assisted cooling.
Processing pentabytes of data
Very high
Uses hundreds of thousands of GPUs, which can pull up to 1.2 kW per chip 150-230 kW per rack
Air cooling is ruled out by the high heat, so you need direct-to-chip liquid cooling
Running a trained model to answer prompts
Moderate per request, but massive at scale
10 to 50 basic queries can evaporate 500 mL of water
Water usage is 9L / kWh
Images take 5-30 times more energy (about as much as charging a smartphone from 0 to 100%). And 20-100 times more water (about 500 mL)
5 sec of video generation can take a further 30-150 times more energy (running a dishwasher for a cycle)
The free tier of something ChatGPT, Claude, or Gemini puts you around 10-80 Wh / day (about as much as a fridge uses running for 1 hour)
Weather forecasting, scientific research, cryptography
High bursts of energy
High water usage
24/7 baseload demand Long-duration loads
Traditional evaporative cooling systems cost millions of gallons of water / year
Use water as refrigerant but lithium bromide as the absorbent
Data centers are meant to couple into business or consumer-use infrastructure. This makes them bunch together, leading to heavy loads on the power grids in those locations.
Northern Virginia has the world's largest concentration of data centers
Large, creditworthy Willing to contract for firm capacity Tolerance for on-site or behind-the-meter solutions