The bottleneck moved to the power grid

The eye-watering numbers in AI this month weren't benchmark scores. They were measured in gigawatts. OpenAI and NVIDIA announced a partnership to deploy 10 gigawatts of systems, with NVIDIA investing up to $100 billion as each gigawatt comes online. Days around it, OpenAI and AMD committed to 6 more gigawatts. And a Stargate data center in Michigan got approved to draw 1.4 gigawatts — over the loud objections of residents who weren't given a hearing.

A gigawatt is roughly the output of a full-size nuclear reactor, or the draw of a mid-size city. The frontier labs are now negotiating for cities worth of power. That tells you where the real constraint has moved.

The bottleneck isn't intelligence anymore

For a while the limiting reagent in AI was model capability. That's not the wall now. The wall is electricity, land, cooling, and the local politics of getting them. The models are ready to do more than the grid can power.

This is why the deals are about energy, not algorithms, and why the friction is now showing up at the level of a township utility commission instead of a research lab. When the binding constraint is physical infrastructure, the game changes from "who has the smartest model" to "who can actually power it" — and the costs and the backlash land on real places with real people.

Efficiency is a moat again

Here's the part that matters even if you'll never build a data center. When power is the scarce resource, every token you don't need to compute is worth real money and real watts. The era of "throw the biggest model at everything" is colliding with a physical bill.

That reframes the boring engineering choices as strategic ones:

• Right-size the model. Routing the easy 80% to a smaller, cheaper model isn't penny-pinching; it's the difference between a workload that scales and one that doesn't.
• Stop re-computing what didn't change. Caching, deltas, and not re-sending the world every call are now economic decisions, not just tidy ones.
• Measure cost per request like you mean it. When the constraint is physical, the efficient design wins the way the clever one used to.

I learned this at a much smaller scale building a real-time system for ten thousand players in one world: the bottleneck was never the CPU, it was the traffic — the bytes you push. Name the real constraint and you optimize the right thing. At civilizational scale, the constraint is the grid, and the same discipline applies — the cheapest watt is the one you never spend.

The bottleneck is physical, so the advantage is too

The frontier is now gated by things you can't fix with a clever prompt: substations, transmission lines, water, permits, and neighbors who'd rather not host a data center. That's a slower, harder world than software is used to — and it rewards whoever wastes the least.

The bottom line

When the headline deals are denominated in gigawatts and fought over at the county level, the model is no longer the scarce part of AI.

The bottleneck moved from intelligence to electricity — which makes efficiency a competitive advantage again, from the data center down to your per-request token bill. Find the real constraint, then stop spending on what doesn't move it. At this scale, that constraint hums at sixty hertz.