It’s the energy, stupid

Last week I wrote that what humans will do about AI is a problem of economics, not philosophy, and that economics is what I need to ramp up on [Link].

So I started, and I’m jotting down my thoughts as I go, starting with:

If we want to pace AI, energy is probably the only bottleneck we can actually use.

Not chips (design, machine building, or manufacturing). Not data (organic or synthetic). Not some treaty about what labs are allowed to train. Not fucking Accenture.

Just good ol’.. electricity.

Why can’t the labs slow down?

In September, Dario Amodei asked the industry to pace the frontier. Ten days later, Anthropic shipped a model that beat its own flagship on agentic benchmarks, and ninety minutes after that, OpenAI shipped the successor to GPT-5.6: more powerful, at half the price.

That’s not hypocrisy. It’s game theory. Dixit and Nalebuff, on the simplest rule there is:

When a simultaneous-move game has this special feature, namely that for a player the best choice is the same regardless of what the other player or players choose, it greatly simplifies the players’ thinking and the game theorists’ analysis. [..] The name given by game theorists for this property is dominant strategy. [..] RULE 2: If you have a dominant strategy, use it.

Dixit, Avinash K.; Nalebuff, Barry J.. The Art of Strategy. Kindle Edition.

Racing is every lab’s dominant strategy. If the others slow down, you race and win. If the others race, you race so you don’t lose. Either way, you race.

And that’s the prisoners’ dilemma: everyone plays their dominant strategy, and everyone ends up worse off than if they had all cooperated. Every lab is defecting, and none of them can afford to stop alone.

The book’s answer is the usual one. You get players out of the dilemma either by rewarding cooperation or by punishing defection. Rewards are tricky, because a player can pocket the reward and defect anyway. Punishment works, but only if you can see the cheating and actually make it hurt.

So the question becomes: what can we see, and what can we make hurt?

What can we actually control?

Here’s what I don’t think works.

Chips. We’ve been trying export controls for years. Chips get smuggled, rerouted, rented through cloud providers in third countries. They’re small, they’re valuable, and there are millions of them. And it’s hard to regulate them within the US.

Data. It’s everywhere. It’s on the internet, it’s synthetic, it’s generated by the models themselves. It’s hard to measure, and hard to keep from being transferred.

GPUs, TPUs, Trainiums. You can’t meaningfully regulate who installs a rack or what job it runs. Once the hardware is in a building, what it does is invisible from the outside.

Now look at energy. A frontier training run or a fleet of inference clusters needs a power plant’s worth of electricity. Power plants can’t hide. Grid connections are permitted, metered, and billed. Even the off-grid plants need air permits, and you can count their turbines. Energy is already one of the most regulated things in the economy! We don’t need to invent an institution to measure it; utilities have been doing it for over a century.

Elinor Ostrom’s work on how communities manage shared resources like fisheries and forests without destroying them earned her a Nobel in Economics. There’s worse company to keep than a Nobel laureate.

Sometimes the rules on what is permissible must be designed in the light of feasible methods of detection. For example, the size of a fisherman’s catch is often difficult to monitor exactly and difficult even for a well-intentioned fisherman to control exactly. Therefore rules based on fish quantity quotas are rarely used. Quantity quotas perform better when quantities are more easily and accurately observable, as in the case of water supplied from storage and harvesting of forest products.

Dixit, Avinash K.; Nalebuff, Barry J.. The Art of Strategy. Kindle Edition.

Chips and data are the fish catch. Energy is the water from storage.

The labs are already scrambling for it

If you want to know what the binding constraint is, watch where the labs are desperate.

Data center developers are now building their own power plants to get around the grid. Behind-the-meter power, generated on-site and never touching the grid, is more than 25% of all planned data center capacity in the US. Only about 2% of that is operating today [Link]. Most of it is gas.

The poster child is xAI in Memphis. They ran 35 gas turbines, and were eventually permitted for 15. In January, the EPA ruled that they had operated the extras illegally [Link]. In April, the NAACP sued over another 27 unpermitted turbines across the state line in Southaven, Mississippi, powering Colossus 2 [Link]. In June, the Department of Justice asked the court to dismiss the case. Its argument wasn’t that the turbines had permits. It was national security [Link]. Dominant strategies, remember?

I find that story encouraging, in a weird way.

Labs don’t break the rules over things that don’t matter. Nobody builds an illegal power plant in South Memphis unless energy is the thing standing between them and the frontier. And the government jumping in to protect it tells you the same thing from the other side.

It also shows that energy is the one place where cheating is visible. You can count turbines from a satellite. People sue over the smog. Nobody has ever sued a lab over their system card, but they’ll sue over the air their kids breathe.

As they should.

Energy quotas, by human seats

So here’s the idea.

Cap the power each lab can use, and set the cap based on its market share of user seats. Humans, actually using the thing.

Big labs are the focus here. A minimal floor gives new labs enough power to get started, and nobody’s building the extinction of humanity on almost no energy anyway.

The quota does three jobs at once.

First, it paces the frontier. Not stops, paces. A lab can’t grow its power budget by deciding to train a bigger model. It grows when more people use what it’s already built. Labs will still train new models; they just have to win more humans to get more power. And we get to define how much of that power goes to serving those humans, and how much to training.

Second, it aligns with keeping humans in the economy by disincentivizing AI that runs without anyone using it. Under a seat-based quota, autonomous agents don’t earn you any extra power. Humans do. The lab’s incentive becomes getting more people into its market share, not taking them out of it, and out of the economy to boot.

Third, it changes what labs spend on. Today, the money goes to research: rushing to train the next model before the competition does. Under a seat-based quota, power spent on training is power you can’t sell to users. Inference becomes the focus of spend, not research.

Which means instead of new models every two weeks, we’d live with Astra and Opus 5.5, except with faster and cheaper inference as time goes on.

Sounds fine to me.

Alignment and control as a commons

Right now, alignment and control research happens mostly inside the labs, funded by the labs, published when the labs want (and however selectively they want), and evaluated by consulting companies the labs pay, like.. you know who.

I think alignment and control research should move out of the labs into a public body, where the research itself is open.

Open, no exceptions. Every finding, every eval, every failed control method that comes out of the public body gets published.

The labs fund it in proportion to their market share. Basically, the money they spend on private control research gets redirected into a shared one. They govern it together through a board, with seats weighted by market share and energy allocation.

The rules apply to the public body, not to the companies. Outside it, private companies can do whatever they want. They’re bound by their energy quota, and that’s it. If they want to spend it mining bitcoin, that’s their problem for all I care.

But a lab that breaks the public body’s rules, say, by taking its research private, or by steering the shared work toward frontier capabilities instead of alignment and control, is cheating. And it gets punished like any other cheater: with a lower energy quota.

Ostrom studied a lot of attempts to manage commons, some successful and some not, and found a handful of prerequisites for cooperation. Clear rules for who’s a member. Clear rules for what’s allowed and what’s forbidden. A system of penalties everyone understands.

An important principle is graduation. The first instance of suspected cheating is most commonly met simply by a direct approach to the violator and a request to resolve the problem. The fines for a first or second offense are low and are ratcheted up only if the infractions persist or get more blatant and serious.

And on detection:

Fourth, a good system to detect cheating must be in place. The best method is to make detection automatic in the course of the players’ normal routine.

Dixit, Avinash K.; Nalebuff, Barry J.. The Art of Strategy. Kindle Edition.

Detection that’s automatic in the normal routine is an electricity meter. Nobody has to hire guards. The utility is already counting.

Her last prerequisite is that the players design the system themselves, because they know the resource and the cheating better than anyone, and “centralized or top-down management has been demonstrated to get many of these things wrong.” That’s why the board is made of the labs. Not a regulator dictating from outside, and not a consultancy on retainer. The labs pay for it and run it together, and none of them controls it alone.

The side effects are (mostly) good

This approach helps with climate change. Most of those behind-the-meter plants burn gas, by the way. Yuck.

It rides popular support that already exists. People don’t want data centers in their backyard, and they don’t need an alignment argument to feel that way.

And it creates an economic incentive for energy-efficient AI. If power is capped, the lab that does more with each kilowatt-hour wins. Sure, efficiency means more capability over time. But most of it goes into inference: today’s models, cheaper and faster, for more people. That’s little incremental risk to alignment and control compared to where we are now. That race, the one for efficiency instead of raw capability, is a much better race to be in.

Compare that to most AI regulation proposals, which ask some group to give up something they want for the sake of safety. Here, the neighbors get cleaner air, the climate gets fewer gas turbines, and the labs give up some power in exchange for a race they can run without betting your kids’ lives on it.

What I don’t know

A lot. Basically.. everything.

People will try to fake seats. I think that’s an auditing problem like any other regulated number, the way companies get audited on revenue, but I haven’t thought it through.

Labs could move to countries with cheap, unregulated power. Which brings in China, which has a lot of the energy the US doesn’t. There’s an interesting version of this where energy becomes a trade, chips and models one way and power the other way. I’m looking into this, but it’ll be a separate article.

And nobody has the authority to set up the board today. I don’t know who would. The US government is both behind and captured, so that’s not it.

Of course I’m not claiming I solved AI pacing in a week. I WON’T SOLVE IT EVER. This is a status update from someone at the very beginning of learning a field, writing down the first things I’m looking at as I figure out what to do.

What I’m doing about it

Studying, for now. I need to understand the field before I know where I can actually act.

Next up, after I get through all this mind-bending game theory stuff over the next few weeks, are two of Vaclav Smil’s books: Energy and Civilization, and Energy: A Beginner’s Guide.. which, by the way, I’ve already read, and it’s certainly the hardest beginner’s guide you’ll ever find. For anything. Ever.

See you next time.