AI is overheating the grid and forcing a cooling revolution in data centers

If your favorite AI app feels slower or more expensive next year, the bottleneck might not be chips. It might be electricity and heat.

Power has officially overtaken land, fiber, and tax breaks as the top factor in where new data centers get built. According to the U.S. Department of Energy’s 2024 Report on U.S. Data Center Energy Use, national data center electricity demand has tripled since 2014 and is projected to double or even triple again by 2028. AI and accelerated computing are the main drivers.

Here is the catch. Every watt powering AI servers turns into a watt of heat. And AI racks are now pushing past 100 kW, with projections of 300 kW later in the decade. Traditional air cooling cannot handle that load.

That is why direct to chip liquid cooling, or D2C, has gone from niche to standard. What once looked risky, with water running through racks packed with expensive electronics, is now mainstream. D2C enables higher sustained clock speeds, better silicon reliability, and denser deployments without runaway fan and chiller costs.

But operators now face a fork in the road. Designs that maximize raw performance are not the same as those that maximize energy efficiency. Some data centers will push for higher TDPs and clock rates. Others will optimize for cost per inference and sustainability. Cooling architecture has become a strategic choice.

AI is even being used to manage the heat. AI assisted cooling controls can predict hotspots, optimize coolant flow, adjust temperature setpoints, and detect early equipment failures. Thermal aware schedulers may soon decide where workloads run based on cooling constraints.

Next up is a battle among advanced systems such as two phase direct to chip cooling, in chip microfluidic cooling, and single phase immersion cooling. The bet here is on two phase direct to chip cooling as the likely standard for extreme heat flux.

In 2026, liquid cooling becomes the baseline. The real question is who can scale AI without burning through the grid or their margins.