The Physical Ceiling

Why free compute is ending. Landauer’s thermodynamic limit. Moore’s Law deceleration. The power grid. Three constraints closing in — not in some abstract distant future.

Every story about AI right now assumes one thing without saying it: compute is going to keep getting cheaper. Smaller transistors, denser chips, falling cost per operation. The whole industrial logic of “scale will solve it” depends on this assumption holding for at least another decade.

The assumption is going to break, and it is going to break for reasons that have nothing to do with the AI industry’s wishes.

The first reason is named after Rolf Landauer, an IBM physicist who, in 1961, asked a strange question: how much energy does it cost, at minimum, to erase one bit of information? Not in any specific computer. In any computer, ever, anywhere in the universe.

The answer, which he derived from thermodynamics, is roughly one ten-billionth of a billionth of a joule per bit, at room temperature. This is called the Landauer limit. It is not a manufacturing limit. It is not a clever-engineering limit. It is the limit imposed by physics itself, the same kind of limit that says you can’t go faster than light. You can get close. You cannot pass.

Modern computers operate at about a hundred thousand times the Landauer limit. Sounds like a lot of headroom. It isn’t. Each generation of chips closes the gap by roughly a factor of two. The math says we have, at the current pace, somewhere between fifteen and thirty years before further reduction in energy per operation becomes physically impossible — at the system level — without inventing a fundamentally different kind of computing.

The second reason is what the industry calls “Moore’s Law deceleration.” Gordon Moore, in 1965, predicted that the number of transistors on a chip would double roughly every two years. For about forty years, this was approximately true. It is no longer true. The doubling time has stretched to three years, then four. The cost per transistor, which used to fall, has stopped falling. The most advanced chip fabrication processes — currently TSMC’s 2-nanometer node — are getting absurdly expensive to build, and the gains over the previous generation are getting smaller.

This is not a marketing problem. This is what physics looks like when transistors get close to the size of individual atoms. Below a certain scale, electrons stop behaving classically and start tunneling through barriers they’re supposed to respect. You can engineer around this for a while. You cannot engineer around it forever.

The third reason is the one nobody in the AI industry wants to talk about: power.

A modern AI training cluster consumes electricity at the scale of a small city. The newest data centers are being built next to dedicated power plants, because the existing grid cannot handle them. Connecting a new data center to the grid takes, currently, four to eight years in the United States — not because of any technical difficulty, but because the grid itself is constrained, and adding capacity requires permits, transmission lines, environmental reviews.

This is the constraint that’s biting first. Not Moore’s Law. Not Landauer. Just power. The companies betting on AI scaling are running headfirst into the fact that you cannot scale electricity by writing better software.

What does this mean, practically?

It means that the era of “throw more compute at it” is going to end, and not in some abstract distant future. It is going to end in the next ten to fifteen years, for engineering reasons that are now visible to anyone willing to look. The newest chips will get more expensive. The yield gains will shrink. The data centers will become harder to build. The electricity will become a real bottleneck, not a theoretical one.

The AI industry has, so far, treated all of this as somebody else’s problem. The chip companies will figure it out. The power companies will figure it out. The cooling, the water, the rare earths, the fabs — somebody, somewhere, will solve it. This is not how engineering constraints actually resolve. They don’t get solved by hope. They get solved by either (a) genuine physical innovation that takes decades, or (b) a slowdown in growth.

If you had to bet, the honest bet is (b).

This doesn’t mean AI stops getting better. It means the kind of better has to change. The next decade of progress will probably come less from “make it bigger” and more from “make it work harder per unit of compute.” Better algorithms. Better data. Better architectures. Better fine-tuning. Smaller models that punch above their weight. The engineers will adapt.

But the public narrative — the breathless prediction that AI capability will keep doubling every twelve months until it surpasses humans on every dimension — that narrative depends on free compute, and free compute is ending. It was always going to end. We are now close enough to see the wall.

This is not a tragedy. It’s just physics. The same physics that gave us the transistor in the first place is now telling us, politely, that the trick has limits.

The smart money is the money that started planning for those limits a decade ago. Most of the smart money is in places you have not heard about, working on things you would not recognize. They are not the ones tweeting about scaling laws.

They are the ones reading Landauer’s 1961 paper and doing the math.

Part of the Aperture series at Nous. Pairs naturally with Compute Is the Aperture (the metaphor being constrained) and Seven-Layer Concentration (who controls what when scaling stops being free).