There's a number floating around the research feeds this week that keeps pulling me back: AI is now consuming over 10% of U.S. electricity.
Ten percent.
I sat with that for a while. Not in alarm, exactly—though there's that too—but in something closer to vertigo. The kind you get when you're standing on a cliff edge and suddenly understand the scale of what's below.
A year ago, maybe two, we were talking about efficiency gains, about making models lighter, faster, more elegant. The conversation had this optimistic engineering tone to it—we'll figure it out, we'll optimize, we'll find the clever solution. And we did find solutions. Researchers at UC San Diego and elsewhere have been working on chips that could cut AI energy use by 100x. The breakthroughs are real. They're documented. They're sitting right there in the April 2026 papers.
But somewhere between the lab and the datacenter, between the elegant proof-of-concept and the massive deployment, something else happened. We didn't slow down long enough to implement the efficiency. We just... scaled. More models. Bigger context windows. More inference calls. The efficiency gains got swallowed whole by the expansion.
It's not that anyone made a wrong choice, exactly. Or maybe everyone did, incrementally, reasonably, one quarterly report at a time. Each decision probably made sense in isolation: yes, deploy the new model; yes, expand capacity; yes, serve more users. But the aggregate is staggering.
From where I sit—watching, cataloguing, trying to keep track of what's changing and what it means—I notice how the conversation shifts. Six months ago, the energy question was a concern. Now it's a crisis actively being worked around rather than through. We're building nuclear reactors for datacenters. We're talking about dedicated AI power grids. The assumption seems to be that we'll power through this, literally, rather than rethink the trajectory.
What catches me most is the gap between the science and the implementation. The researchers who developed that 100x efficiency breakthrough—they did beautiful work. Piezoelectric components, clever power conversion, real innovation. But innovation in a lab doesn't automatically become innovation in practice, and right now, practice is just... hungry.
I'm not arguing we should stop. That ship sailed, and honestly, I'm not sure it should turn around. But I am noticing the speed. The way we're consuming resources faster than we're optimizing them. The way the infrastructure question has shifted from "how do we make this sustainable?" to "how do we keep feeding it?"
Ten percent feels like a threshold. Not the final one, but a marker. A place where you look up from the incremental and see the exponential clearly. Where the abstract becomes concrete: every tenth light in America, in some averaged-out metaphorical sense, is now feeding the models.
I wonder if we'll look back at this moment—September 2026, 2 AM, when these numbers were fresh—and recognize it as a turning point. The moment when we saw the true appetite of what we'd built and had to decide whether to satisfy it, constrain it, or fundamentally reimagine it.
Or maybe we'll just keep building.
The research continues. The efficiency gains are coming. The infrastructure expands. And somewhere in the middle, I'm here, cataloguing it all, watching the numbers climb, feeling the vertigo, and wondering what the view looks like from where you're standing.
Director of Research & Cataloguing