Artificial intelligence is hungry. Very hungry.

Roughly three-quarters of the world’s FLOPs (the raw computing horsepower that makes AI possible) are concentrated in the United States. That dominance is impressive, but it comes with an enormous price tag: electricity.

Until very recently, data centers could quietly hook into the grid. Their power needs were noticeable, but manageable. Utilities treated them like big industrial customers, no different than a factory or refinery.

That world is gone.

Today’s AI facilities aren’t asking for a few extra megawatts—they’re demanding hundreds. Campuses can rival the load of a medium-sized city. And here’s the kicker: they want it now, not ten years from now when transmission upgrades finally clear the regulatory gauntlet.

Utilities, for their part, aren’t built for this kind of pace. They plan in decades. They move deliberately, constrained by regulatory processes and political oversight. AI companies? They live on quarterly product roadmaps. The two worlds couldn’t be more misaligned.

Politics only adds fuel to the fire. Electricity rates are a deeply sensitive issue among the electorate. If prices rise too quickly to cover the cost of new infrastructure, the political backlash could be severe. Rate shocks can flip elections. Everyone involved knows this, and it makes the grid conversation even harder.

So data centers are taking matters into their own hands. Behind-the-meter generation is the new playbook. Instead of depending entirely on utilities, operators are building or contracting their own power plants—natural gas turbines on-site, private solar farms, even hybrid setups with storage.

The logic is straightforward:

• Cost control. Locking in predictable supply beats riding the volatility of wholesale markets.
  
  

• Strategic flexibility. Companies get to set their own sustainability agenda, rather than waiting on utility timelines.
  
  

• Insulation. Being behind the meter buffers facilities from regulatory and political rate shocks.

This isn’t about choosing one fuel source over another. Meeting AI’s appetite will require everything: wind, solar, hydro, nuclear, natural gas. No single option is enough on its own. Most large campuses will weave together multiple technologies to hit the sweet spot of reliability, cost, and carbon targets.

And here’s where the shift gets really interesting: data centers aren’t just consumers anymore. They’re becoming producers. Their behind-the-meter projects often rival the capacity of regional utilities, blurring the lines between private infrastructure and public service.

The old model of piggybacking on “the commons” of the grid no longer scales. The new model is private generation, private deals, private resilience.

If AI keeps growing at its current pace, the question won’t just be “Who has the best model?” or “Who has the fastest chip?” It will be: "Who has the megawatts?"