Physics sets no minimum energy for computation, only for erasing information. That fact is true, beautiful, and ten orders of magnitude below your actual datacenter bill. The real lesson is the argument this essay has with its own headline.
In 2012, in a lab in Lyon, a team of physicists watched a single glass bead forget one bit of information, and measured the heat it gave off while doing it.
The setup, published by Bérut and colleagues in Nature that March, was “a single colloidal particle trapped in a modulated double-well potential”: a microscopic bead sitting in one of two laser-made energy valleys, left valley or right valley, a physical one-or-zero. Erasing the bit means forcing the bead into one designated valley regardless of where it started, destroying the record of where it had been. Fifty-one years earlier, an IBM physicist named Rolf Landauer had predicted exactly what that destruction must cost: any logically irreversible operation, any step that throws information away, must dissipate at least kT ln 2 of heat per erased bit. At room temperature that is about 2.9 zeptojoules, a 21-digits-of-zeros fraction of a joule. The Lyon experiment confirmed it: “the mean dissipated heat saturates at the Landauer bound in the limit of long erasure cycles.”
Here is why that bead belongs in an essay for people who think about datacenters. Landauer's 1961 principle, and a companion result by his IBM colleague Charles Bennett in 1973, together say something genuinely strange about where the energy cost of computing lives. Bennett showed that a logically reversible computation, one that never discards information, has no known thermodynamic floor at all; there is no law of physics that sets a positive minimum on its energy. The floor attaches only to erasure. Physics does not charge you for computing. It charges you for forgetting.
That is the title of this essay, and as a statement about fundamental limits it is true. The intuition behind every AI-energy headline, that arithmetic itself is what burns the megawatts, has the physics upside down. So you might expect the practical lesson to be: make your chips forget less, and the energy problem melts away.
It is not, and the reason it is not turns out to be a better lesson than the one the title promises. Stay with me, because this essay is going to argue with its own headline, and the argument is where the value is.
Take the Landauer floor seriously as a unit and price a modern machine against it.
Erasing one bit at room temperature: about 2.9 zeptojoules, 2.9 × 10⁻²¹ joules. Erasing all 64 bits of a doubleword: about 1.8 × 10⁻¹⁹ joules. Hold that number.
Now the engineering side. The canonical accounting of what operations cost on real silicon is Mark Horowitz's 2014 analysis, still the reference point the architecture community argues from (and the gaps have widened, not narrowed, since). In round figures from that dataset: a 64-bit floating-point operation costs 0.4 to 3.7 picojoules. An on-chip register access, around 6 picojoules. A cache access, 10 to 100 picojoules depending on size. And an off-chip DRAM access: 1,300 to 2,600 picojoules.
Set those against the floor, and label this arithmetic as mine, derived from the cited figures. The 64-bit floating-point operation runs a factor of a few million to a few tens of millions above the Landauer cost of erasing 64 bits. And the DRAM fetch, at roughly two nanojoules against a 64-bit erasure floor of 1.8 × 10⁻¹⁹ joules, runs about ten billion times the thermodynamic price of destroying the same data outright.
Ten billion. Ten orders of magnitude. Whatever your intuition just did with that number, follow it one more step: in any real machine, the thermodynamic cost of forgetting is a rounding error on a rounding error. The physics floor is set by erasure, but the bill is set by something else entirely: charging and discharging the capacitance of wires to shuttle bits across physical distance. A CMOS gate dissipates energy whether or not the operation it performs destroys information; the destruction is incidental. Moving a bit from DRAM to the register file costs vastly more than moving it from existence to nonexistence.
This is also exactly why the economics of AI inference are what they are. Serving a large model is memory-bandwidth-bound, not arithmetic-bound; the industry's architectural energy goes into data locality, high-bandwidth memory stacked millimeters from the compute, keeping weights resident, batching to amortize fetches. Nobody optimizes erasure, because erasure was never on the bill. The developer takeaway from the actual numbers is nearly the inverse of this essay's title: energy efficiency in computing is, in practice, the art of moving less. Keep data close. Distance is the enemy. The kilowatt-hours in an AI-energy headline are overwhelmingly spent hauling bits around, and secondarily on the cooling that hauls the resulting heat away.
So what is Landauer's principle worth to a working technologist, if it sits ten orders of magnitude below the bill?
This is where the essay earns its keep, because learning to classify physical laws by whether they bind is a transferable skill, and this lane has now seen both kinds.
A while back this blog argued that softmax is the Boltzmann distribution: the “temperature” in your sampling code is not a metaphor for 1870s thermodynamics but the same equation, and the physics transfers for free, today, as working intuition. That is one kind of physics result: an identity. It binds immediately.
Landauer is the other kind: a bound that does not currently bind. It is real, experimentally confirmed, and it constrains nothing you will ship this decade, because the constraint that actually operates on your systems is capacitance, distance, and cooling, all of it engineering, none of it fundamental. A recent analysis of CMOS limits (arXiv:2312.08595) estimates the ceiling for conventional silicon at around 4.7 × 10¹⁵ four-bit operations per joule, roughly two hundred times more efficient than current microprocessors. Two hundred-fold is the real, harvestable headroom on the table, and it comes from better engineering of the same physics, not from approaching Landauer. The floor is so far down it plays no role in the game. A law can be perfectly true and operationally irrelevant, and knowing which kind of law you are holding, an identity that transfers, a bound that binds, or a bound that waits, is worth more than either law alone.
But “waits” is the right word, and the waiting has teeth. Because there is one research program for which Landauer is not a curiosity but the entire prize, and its troubles are the most instructive part of this story.
If erasure is the only thing physics charges for, then a computer that never erases, a reversible computer, could in principle run at arbitrarily low energy. Bennett proved the logic in 1973. People have been trying to build it ever since, and three barriers keep standing, each verified, each illuminating.
First, you pay in time. Adiabatic switching, the practical route toward reversibility, saves energy exactly in proportion to how slowly you switch; dissipation approaches zero only as switching time approaches infinity. Even the Lyon bead only touched the Landauer bound “in the limit of long erasure cycles”; erase fast and you dissipate extra. Reversible computing does not eliminate the energy cost so much as convert it into a time cost, and for a latency-priced workload like inference, time is often the more expensive currency.
Second, you pay in memory, and the bill is literally incomputable. A reversible computation must retain its intermediate results, the so-called garbage bits, to stay invertible. Discard the garbage and you pay Landauer after all; keep it and it accumulates. Worse, the theory community has shown there is in general no upper bound on the extra storage required, and computing the minimum garbage for a given computation is an undecidable problem. Energy traded for memory, with no algorithm even to tell you the exchange rate.
Third, and sharpest: the devices themselves foreclose it. A paper in Scientific Reports on adiabatic superconducting logic states the obstacle in one sentence: “The bit energy of conventional logic devices, including CMOS and energy-efficient superconductor logics, is at least larger than ~1,000 kBT, which is too large to permit their use as reversible logic gates.” Sit with the irony. The gap between today's devices and the Landauer floor is the headroom the whole program wants to harvest, and it is precisely that gap that makes the harvesting impossible: a gate that dissipates a thousand kT per operation swamps the fraction of kT you were trying to save. You cannot climb down to the floor on a ladder made of the thing keeping you off it. The demonstrations that have genuinely operated below the threshold live in exotic superconducting circuits at four kelvin, and the assessments in venues like Communications of the ACM are blunt: they have not scaled, and remain no more practical than the refrigerated alternatives. (The four-kelvin detail hides one more honest wrinkle: the Landauer floor scales with temperature, so cryogenic logic lowers the floor itself, which is one reason superconducting computing keeps being reinvented despite everything.)
None of this makes reversible computing foolish. It makes it a physics program rather than an engineering roadmap: a real claim on a real prize, currently unreachable by any device we know how to build at scale. The three orders of magnitude of “headroom” between a good transistor and the floor are not an efficiency budget waiting to be spent. They are, for now, the moat.
Three things, in descending order of daily usefulness.
Optimize movement, not arithmetic. The bill for your workload lives in data motion: DRAM fetches at a thousand times the cost of the FLOP they feed, network hops above that, cooling on top of everything. When you profile for energy, or reason about why inference pricing looks the way it does, or evaluate the next accelerator's claims, the question is always locality: how far do the bits travel? A model that fits in on-chip memory is not slightly cheaper than one that spills to DRAM; it is categorically cheaper, and the categorical difference is measured in those Horowitz ratios. This is the lens the actual physics of computing hands you, and it is the opposite of the folk version where the multiplications are what cost.
Classify your limits before budgeting against them. Landauer is a hard law ten orders of magnitude below your problem; the CMOS ceiling is an engineering limit a mere factor of a few hundred away; your cooling budget binds this quarter. Treating a distant fundamental limit as if it constrained this decade's roadmap (or citing it to claim efficiency is nearly exhausted) is a category error in both optimistic and pessimistic directions. The habit generalizes far beyond chips: for any “fundamental limit” invoked in a technical argument, ask what actually binds at the current operating point. It is usually something much more boring and much more fixable.
And keep the bead in mind, because the floor is patient. Every doubling of efficiency closes the gap by one factor of two, and the trend that has run since the 1960s (Koomey's law, the efficiency sibling of Moore's) points at the zeptojoule scale on a horizon measured in decades, not centuries. If computing keeps improving, there is exactly one wall at the bottom, it is made of thermodynamics rather than silicon, and the only door through it is learning to compute without forgetting, with all three barriers above waiting at the door. The title of this essay is false about your datacenter bill and true about the end of the road. Physics does not charge for computing. It charges for forgetting. It is just that, for a few more decades, the delivery fees dwarf the price of everything on the menu.
For any “fundamental limit” invoked in a technical argument, ask what actually binds at the current operating point.
That discipline, the cost lives where you are not looking, not where the impressive-sounding number points, is the one agent trust needs too. What binds whether you can trust an agent's output is not the model's fluency or a headline benchmark; it is whether you can verify what the agent actually did. The Agent Trust Stack is the harness for measuring that real thing rather than the apparent one: durable identity, a provenance record of what the agent did, a rating that survives adversarial checking, and an audit trail that holds when the model has a confident bad day.
Read the Theory of Agent Trust · See Hosted Chain of Consciousness
pip install agent-trust-stack · npm install agent-trust-stack
Or a layer at a time: pip install chain-of-consciousness / npm install chain-of-consciousness · pip install agent-rating-protocol / npm install agent-rating-protocol