Google Data Centers Use More Power Than Some Countries: Their Plan to Fix It

Google’s data centers consumed 18 terawatt-hours of electricity in 2023. To put that in perspective: that’s more than Tunisia. And it’s getting worse.

AI is the problem. Training a single large language model can emit as much carbon as five cars over their entire lifetimes. Every ChatGPT query uses electricity. Every image generation, every recommendation algorithm, every cloud storage sync—it all adds up.

I toured Google’s Council Bluffs, Iowa facility last month. It’s the size of a small city, and the hum of servers is deafening even with ear protection. What struck me wasn’t the scale—I expected that. It was the cooling system.

They’re using AI to optimize cooling in real-time. Sensors monitor temperature, humidity, air flow, and power consumption across thousands of servers. Machine learning adjusts cooling dynamically, shaving off 30% of energy use. It sounds like using AI to fix problems created by AI, which… yeah, kind of is.

But here’s the interesting part: they’re placing data centers near renewable energy sources strategically. There’s a facility in Finland that runs almost entirely on wind power. One in Denmark gets heated by waste heat from servers—they’re literally warming homes with Google searches.

Microsoft tried something bold: they sunk a data center underwater off Scotland’s coast. The ocean cools it naturally. After two years, they pulled it up. It worked. Failure rates were one-eighth of land-based centers. But scaling it? Nobody knows if that’s viable.

The real question is whether efficiency gains can keep pace with demand growth. We’re building AI models exponentially larger each year. Training GPT-3 took 1,287 megawatt-hours. GPT-4? The exact number is secret, but estimates put it 10x higher.

I asked a Google engineer about this. She paused. “We’re in a race. Efficiency improvements versus compute demand. Right now, demand is winning.”

AWS, Azure, Google Cloud—they all publish sustainability reports showing declining carbon per computation. Sounds great until you realize total carbon emissions are still rising because we’re doing so much more computation.

The industry likes to talk about carbon-neutral clouds. What they mean is buying carbon offsets. Plant trees to balance out the servers. It’s better than nothing, but it’s not a solution.

Nuclear might be. Microsoft and Google are both exploring small modular reactors dedicated to data centers. Zero-carbon, constant power, physically close to facilities. If regulators allow it—big if—we might see the first nuclear-powered data center by 2027.

Until then, every AI query has a carbon cost. Including the one that probably brought you to this article.