The cycle that began with [YOU] on AI describes AI's outputs as appearing frictionless and instantaneous at the user interface. The Landauer limit names the friction that the interface conceals: the dissipation of heat in the physical substrate of every computation, scaling with every increase in capability and every expansion of deployment. The dream of an intelligence that grows without bound, recursively improving itself at no thermodynamic cost, runs straight into this limit. There is no demon that sorts for free.
Practically, the limit's relevance is growing rather than receding. The energy consumed by frontier AI training has become a strategic variable—a constraint on how large the next model can be, where data centers can be built, what the carbon implications of capability growth are. As hardware efficiency improves, the gap between actual consumption and the Landauer floor narrows, making the floor's existence progressively less theoretical. The limit is the long-run wall that scaling approaches asymptotically.
Rolf Landauer published “Irreversibility and Heat Generation in the Computing Process” in the IBM Journal of Research and Development in 1961, working from the thermodynamic tradition that Maxwell and Boltzmann had built. His central insight was to distinguish logically reversible from logically irreversible operations: a computation that maps two possible inputs to a single output destroys information and must therefore dissipate entropy. Leó Szilárd had proposed in 1929 that the thermodynamic price was in the demon's act of measurement, but Landauer showed measurement could in principle be reversed; it was erasure that was necessarily irreversible.
Charles Bennett completed the resolution of Maxwell's demon in the 1980s, showing that the demon could operate indefinitely by running logically reversible measurements and accumulating a record, but that at some point it must erase to free memory for more sorting, and each erasure repays the entropy the sorting seemed to have stolen. The accounting always balances; the second law survives. Landauer's principle was confirmed experimentally in 2012 by Éric Lutz and colleagues at a scale one-fifteenth above the theoretical minimum.
Irreversibility as the source of heat. It is not computation per se that generates heat but logically irreversible computation—operations that discard information. A computation that can be run backward, recovering its inputs from its outputs, is reversible and approaches the thermodynamic ideal. Most conventional computation is saturated with irreversible operations; most of an AI system's operations are logically irreversible.
The floor AI approaches as it scales. Today's chips waste energy at rates billions of times above the Landauer limit. As hardware efficiency improves and as systems scale to fill the efficiency gains with more computation, the gap narrows. The limit is the asymptote that engineering approaches but cannot cross: a minimum cost per bit discarded that no innovation can eliminate. At the scale of frontier large language model training, even approaches toward the limit would represent enormous savings, making the theoretical limit an engineering target.
Learning as disciplined forgetting. A model learns by compressing a vast corpus into a finite weight matrix—by keeping the signal and discarding the noise, by extracting the pattern and letting the individual instances go. This is forgetting on purpose, and by Landauer's accounting it is forgetting on purpose that costs energy. The intelligence of the system is built from acts of erasure; its generalization is the residue of compression; and the compression pays the Landauer price for every bit discarded.
Reversible computing as the path forward. If erasure is what costs, a computer designed around reversible logic—in which every operation is its own inverse—could in principle approach the thermodynamic ideal. This is the engineering descendant of Maxwell's demon resolution: a sorting machine that never erases and therefore never pays. Whether reversible computing can be made practically competitive with conventional architectures—whether its temporal and spatial overheads defeat its thermodynamic advantages—is an active research question and one of the live engineering implications of a nineteenth-century thought experiment.