Yes, on three fronts, though none is a magic wand yet.
First, cooling without losing water. Newer data centres are switching to closed loop liquid cooling, where the same water circulates through pipes to the chips and back again, like the radiator in a car, instead of evaporating away.
Microsoft’s latest designs use this to run with almost no water consumption, and studies show liquid cooling can cut water use by 30 to 50 percent compared with air cooling.
Second, chips that sip rather than gulp electricity. Neuromorphic chips copy the human brain, which runs on about 20 watts, less than a dim bulb.
IBM’s NorthPole research chip is roughly 25 times more energy efficient than an older NVIDIA GPU and does not need liquid cooling at all.
Even more radical are photonic chips, which calculate using light instead of electric current. Because light does not heat the chip the way electricity does, German firm Q.ANT claims its commercial photonic processor cuts energy use for AI tasks by up to 30 times while needing far less cooling.
Third, plain efficiency. The IEA notes that the power used per AI task is falling at a pace it calls unprecedented in energy history.
The reality check: no chip runs on zero electricity, and none of these newcomers can yet train or run the giant models behind ChatGPT or Gemini. For now, the heavy lifting still happens on power hungry GPUs. The escape routes are real, but they are exits being built while the traffic keeps growing.