One kilogram could out-think every AI chip on Earth

Every step a computer takes means flipping some piece of matter from one state to another, like a switch going from off to on. Quantum mechanics caps how fast anything can flip, and the cap depends on how much energy it has 12.

In 2000 the MIT physicist Seth Lloyd pushed that rule as far as it goes. He imagined a laptop weighing one kilogram, gave it every joule that mass can hold through E = mc² (about 9×10¹⁶ joules), and worked out its top speed: 5.4×10⁵⁰ operations per second 1.

Now compare that with what we've actually built. Epoch AI estimates the world's AI chips add up to about 31.6 million top-of-the-line Nvidia H100s' worth of computing as of September 2026 3, and each H100 manages roughly 2×10¹⁵ operations per second 4. Multiply them out and the whole planet's AI fleet runs at about 6×10²², which leaves 28 powers of ten of empty room above it.

Lloyd wasn't the first to find a ceiling like this. In 1962 the mathematician Hans Bremermann worked out a similar limit of about 1.4×10⁵⁰ bits per second for every kilogram of computer 5. Two different routes landing in the same place is usually a good sign the physics is solid.

Memory has a ceiling too

Speed is half of a computer, and memory is the other half. Its limit comes from entropy, which counts how many distinct arrangements a lump of matter can take, and Lloyd's kilogram could store about 2×10³¹ bits 1. Max Tegmark estimates your synapses hold "roughly 100 terabytes' worth of information" 6, so one ultimate kilogram could hold the memories of about 27 thousand trillion people.

Here's the catch, and it's my favorite part of the paper. To run that fast and use all that memory, the laptop has to turn essentially all of its matter into energy. Lloyd describes the result as "a plasma at a billion degrees Kelvin: the laptop's memory looks like a thermonuclear explosion or a little piece of the Big Bang!" 1

He admits, dryly, that "packaging issues alone make it unlikely that this limit can be obtained" 1. The point of the exercise is to measure how much room exists above us, and the answer is an absurd amount.

Real chips run into a much closer limit. In 1961 Rolf Landauer showed that erasing a single bit of information always releases a tiny puff of heat 7, about 3×10⁻²¹ joules at room temperature. It's a cost no computer that throws information away can avoid, and physicists finally measured it directly in 2012 8.

Today's machines sit far above that floor. The most efficient supercomputer in the world, KAIROS in Toulouse, does about 73 billion operations per watt 9, which works out to roughly 1.4×10⁻¹¹ joules per operation, about five billion times the Landauer cost. That comparison exaggerates the gap a little, since one operation flips many bits, but the room left over is still enormous.

There's even a known way under the floor. Charles Bennett showed in 1973 that a computer which never throws information away could, in principle, spend far less energy per step than Landauer's cost 10, an idea called reversible computing. It has started to leave the blackboard: a startup called Vaire Computing tested a chip in August 2025 that "drew around 30 percent less energy than a regular chip performing the same computation" 11.