The tech industry spent a decade telling you it was going to save the planet. Clean energy, carbon neutral by 2030, all that. And now the same companies are quietly ordering gas turbines like it’s 2005. But why ?
Training AI models and running them for millions of users takes an insane amount of electricity. Data centers worldwide are burning through roughly 500 TWh this year, about 2% of global consumption. That number is doubling by 2030. Meta, Amazon, Google and Microsoft alone are spending 320 billion dollars on AI infrastructure in 2025.
The obvious move is renewables, right? Solar is the cheapest energy source in most of the world, we’re adding 600 TWh of solar capacity every year, problem solved. BUT... we have a boring infrastructure detail that changes everything. You can build a data center in 2 years but building the high-voltage power lines to connect it to a solar farm in the desert takes 7 to 15 years. Permits, land negotiations, environmental reviews, lawsuits. The electricity exists but you just can’t get it where it needs to go.
And we can’t just put solar farms next to data centers because they are located where the fiber optic infrastructure and users are (Virginia, Ohio, major cities), not where the best sunlight is (deserts). Moving a data center away from fiber hubs means higher latency, which means degraded performance.
So what do you do when you have a 2 billion dollar data center sitting there waiting for power and the grid can’t deliver? You buy gas turbines, plant them on site, and plug them and in 12 to 18 months that’s done. That’s exactly what Meta is doing in Louisiana with a 1.5 GW campus running on natural gas.
The irony is genuinely painful. The technology that’s supposed to represent the future is being powered by the energy source we were supposed to be moving away from. And it gets worse when you look at water. A single large data center can drink 5 million gallons of water per day, roughly what a city of 50,000 people uses. But that’s just the direct cooling. The power plants feeding those data centers consume about 10 times more water upstream. In Texas, data centers are projected to use 49 billion gallons this year. The more you lean on gas and coal to power them, the more water you burn through, because unlike solar and wind, thermal plants need massive amounts of water for cooling.
So the real question isn’t “does AI use too much energy.” At 2% of global electricity, it’s manageable in absolute terms. The question is what kind of energy, and the honest answer right now is: a lot of it is fossil fuel, and the infrastructure bottleneck means it’s going to stay that way for a few years.
The solutions exist but they all have timing problems. Solar and batteries are scaling fast but need grid connections that take a decade to build. Small modular nuclear reactors could sit right next to data centers and run 24/7 with zero emissions, but none are commercially operational yet, probably not before 2029. And about nuclear fusion, it’s still in the lab. Models are getting more efficient per query, which helps, but total compute demand is growing faster than efficiency gains.
The most likely scenario is ugly and pragmatic. Gas fills the gap from now until about 2028, solar plus batteries ramp up through 2030, fission finally shows up around 2030-2035 (I’m talking about SMR), and by then the models are efficient enough that the most extreme demand projections never materialize. We get there, but we take a dirty detour to do it.
The thing that frustrates me about how this gets discussed is that it’s always framed as either “AI is destroying the planet” or “it’s fine, renewables will handle it.” Both are wrong. The planet can produce enough clean energy for AI. It just can’t build the wires fast enough and until it does, the gap gets filled with gas.
That’s not an AI problem. It’s an infrastructure problem that AI is exposing. And if we’re being honest, it’s been there for decades. We just didn’t have a reason to care until now.