The Thermodynamic Cost of Answering

1. The missing ceiling

From 1946 to 2009, computing efficiency doubled every 1.5 years, and workloads moved from mainframes to desktops to pockets when the efficiency gains made local computation sufficient. The same transition is now underway for AI, and recent measurement work has proposed intelligence per watt, task accuracy per unit power, as the metric to track it [1].

Intelligence per watt, as it stands, compares artifacts to artifacts. Engine efficiency had the same character before 1824. Carnot changed the situation by supplying a ceiling fixed by physics, so that the efficiency of an engine became a distance from a law rather than a ranking among machines. This article states and proves the corresponding ceiling for machines that answer questions.

The physical ingredient is Landauer's principle: erasing one bit into a bath at temperature \(T\) costs at least \(k_BT\ln 2\) of dissipated heat, about \(2.87\times10^{-21}\) joules at room temperature. The information-theoretic ingredient is rate–distortion theory. The result is a bound of the form

per query, where \(R_D(a)\) depends only on the task and the required accuracy, together with a precise account of when the bound holds, two explicit machines that evade it when its conditions are dropped, and a protocol that attains it.

2. The intelligence rate

A task is a query alphabet with distribution \(D\), an answer alphabet, and a score \(s(q,\hat y)\in[0,1]\). Abstention, if permitted, is an explicit answer symbol scored by \(s\); accuracy is always the unconditional average over all queries. For an accuracy level \(a\), define

This is the classical rate–distortion function of the source \(D\) with distortion \(1-s\), evaluated at distortion \(1-a\). It measures the minimum number of bits about the query that any answering system must acquire and act on to reach accuracy \(a\). It is computable by the Blahut–Arimoto algorithm, and it is zero whenever a constant answer meets the accuracy level: a task that does not require reading the question has no thermodynamic floor.

Two computed examples, at the accuracy \(a=0.887\) that local models reach on single-turn queries [1]: a uniform binary task has \(R=1-h(0.113)\approx 0.491\) bits, and a uniform four-option multiple-choice task has \(R=2-h(0.113)-0.113\log_2 3\approx 1.312\) bits.

A property worth stating separately: \(R_D(a)\) counts task bits, not text bits. In a numerical check, a task whose answer space carried seven bits of text, of which four were score-irrelevant style dimensions, had \(R_D(a)\) exactly equal to the three-bit style-free value, to machine precision, without the style dimensions being identified by hand. The optimization discards what the score does not see. This is what makes the denominator of an efficiency metric well-defined for open-ended tasks.

3. The bound

The machine is a classical Markov jump process coupled to a single memoryless heat bath at temperature \(T\), with local detailed balance. Internal randomness is part of the machine's state. One operating cycle has three phases: a read phase, during which a register holding the query \(Q\sim D\) is present and readable; an emit phase, after the query register has departed, during which the answer \(\hat Y\) is written to a blank output register that then departs; and a reset phase. The machine operates in steady state, and its accuracy satisfies \(\mathbb E[s(Q,\hat Y)]\ge a\).

Two interface conditions carry the physics.

One-way input. The query register is read-only, departs after the read phase, and is never presented again. The user keeps their query; the machine holds a copy it can never erase against the original.

Copy-semantics output. When the answer departs, the machine still holds a record determining what it emitted: \(H(\hat Y\mid A)=0\) at delivery. Every signal-based channel has this property, because transmitting modulates a field and leaves the sender's registers intact.

Theorem. Under these conditions, the average heat dissipated per cycle satisfies \(\langle Q_{\mathrm{diss}}\rangle\ge k_BT\ln2\,R_D(a)\).

The proof runs the second law through each phase with the finest admissible conditioning: conditional entropies while the query is present, marginal entropies after it departs, since a reset protocol can only implement rates that depend on what the machine can access, and erasure without the reference costs the marginal entropy [5]. The three phase inequalities telescope to

the mutual information between the machine's state at the end of the read phase and the query. The emit phase cannot consult the departed query, so \(Q\to A(t_1)\to\hat Y\) is a Markov chain, and data processing gives \(I(A(t_1);Q)\ge I(Q;\hat Y)\). The final step costs nothing: the machine's induced channel meets the accuracy constraint, so \(I(Q;\hat Y)\ge R_D(a)\) by the definition of the infimum. No coding theorem is used in the converse.

The physical reading is compact. An answering machine manufactures correlation between question and answer, at least \(R_D(a)\) bits of it per query, and in steady state it must destroy its own share of that correlation every cycle. Destroyed correlation is entropy production. A Maxwell demon converts correlation into work at \(k_BT\ln2\) per bit; an answering machine converts work into delivered correlation at the same exchange rate.

4. Two ways to answer for free

Each interface condition is necessary, and each failure mode is an explicit machine.

Uncomputation. Suppose the query register remains readable during the reset phase. The machine computes \(\hat y=f(q)\) reversibly, copies it out, then runs the computation backwards against the still-present query, restoring every internal register to blank. No erasure without a reference occurs, and the heat vanishes in the quasi-static limit. Read access alone suffices. This regime is physical: clamped quasi-static analog networks realize inference with vanishing minimal energy cost [8]. Deployed inference is the opposite regime.

The swap machine. Suppose the machine may deliver its sole physical copy of the answer, move semantics rather than copy semantics. It computes \(f(q)\) into blank memory while the query is present, at no cost, then swaps that register with the blank output register, a reversible permutation, and ends the cycle already blank. Nothing is ever erased and the heat vanishes, at the best deterministic accuracy, for any task. Handing over your only copy destroys no correlation, so Landauer charges nothing; the cost falls on whoever eventually erases the delivered records.

The theorem's scope is therefore exact: answering is thermodynamically costly precisely when the question departs and the answer leaves by copy. Wires, radio, light, and speech all copy. A courier handing over the single written page swaps. Since machines answer by signalling, the bound applies to them unconditionally, and a streaming variant of the proof covers token-by-token emission with retained context, which is how language models are actually served.

5. The price is exact

The frontier is attained. The optimal machine records only a minimal sketch of the query, of size close to \(R_D(a)\), computes reversibly, delivers by copy, uncomputes its scratch against the sketch, and erases the sketch conditioned on internal common randomness that is never modified and never erased. By the strong functional representation lemma [6], a one-shot machine achieves heat at most \(k_BT\ln2\,(R+\log_2(R+1)+4)\), and machines that read blocks of \(n\) queries at once achieve \(k_BT\ln2\,R_D(a)\) per query as \(n\to\infty\), quasi-statically.

The pairing of the converse with this protocol sharpens the engineering statement: reversible computing can eliminate every joule of dissipation above \(k_BT\ln2\,R_D(a)\), and not one joule below it.

6. The efficiency scale

Define the dimensionless intelligence efficiency of a system spending free energy \(\langle J\rangle\) per query at accuracy \(a\):

For chemically driven multi-reservoir systems such as brains, the same proof bounds total entropy production, and \(\eta\) against free-energy consumption is the correct form. Anchors, for a single-turn email-drafting task with \(R_D(a)\sim10^3\) bits:

Machines passed brain parity on commodity single-turn tasks and now lead by roughly two orders of magnitude there, while on long-horizon frontier reasoning the 20-watt organ still leads. The crossover between the two is an iso-\(\eta\) curve moving through task space, and its position is measurable.

The distance to the frontier factorizes, with every intermediate quantity explicit so the numbers multiply exactly. The laptop anchor executes about \(4.8\times10^{12}\) floating-point operations per query; at roughly 32 overwritten register bits per multiply-accumulate this is \(1.5\times10^{14}\) bit-erasures for \(10^3\) task bits, and

Eleven orders of magnitude of headroom are algorithmic, in the number of erasures spent per task-relevant bit, and about eight and a half are hardware, in the system-level energy of each erasure. Koomey's law was the story of the second factor. The first factor is larger.

7. Answering fast

Attainment is quasi-static, so the frontier as stated is reached only at zero power. A finite-time extension bounds the cost of speed. Building \(R\) bits of correlation forces the machine's joint distribution with the query to move a total-variation distance of at least \(\ell^\ast(R)\approx R/(2\log_2|\Omega|)\), and the classical speed limit of Shiraishi, Funo, and Saito [7] converts displacement into entropy production. The result:

where \(N\) is the expected number of state transitions spent reading. Excess heat times operations per query is bounded below by a task-intrinsic constant; latency trades against excess dissipation as \(1/\tau\) at fixed activity, and parallelism relaxes the tradeoff. For macroscopic state spaces the constant is numerically tiny, so the content of the bound is structural: the frontier is a curve in the energy–latency plane, approached only as operations per query grow without bound.

8. Familiar queries

For correlated query streams the proven floor weakens to \(k_BT\ln2\,[R_D(a)-I(A_0;Q)]_+\), where \(I(A_0;Q)\) is the information the machine already holds about the incoming query. A machine that knows its user faces a weaker floor per query. The discount has a deployment consequence: multi-turn context echoes the machine's previous answers back inside the next query, which makes \(I(A_0;Q)\) large, so the full \(R_D(a)\) floor applies to single-turn, fresh-query operation, and the anchors above are defined accordingly.

9. Relation to prior work

The tools are classical: Landauer's principle and Bennett's analysis of the demon, the second law with mutual information [4], erasure against side information [5], and the thermodynamics of prediction [3].

On priority: a novelty sweep run after the derivation was complete surfaced a self-published preprint of Vityaz [9] which states, for the specialized task of reconstructing an external clock from a reusable record, the same converse composition one month earlier: reset work bounded below by an indirect rate–distortion function, by Landauer plus data processing, with a no-side-information assumption. The clock-specialized composition is his. The present work treats arbitrary tasks, derives the cycle bound from phase-wise second-law inequalities rather than assuming the erasure charge, proves both interface conditions necessary through the two zero-dissipation machines above where the earlier work excludes them by assumption, supplies the one-shot attainability constant, and adds the finite-time frontier. Among journal literature the closest work is Still [2], where an information-bottleneck object appears as the dissipation-minimizing representation rather than as a task-indexed floor. Recent bits-per-joule proposals [10] note the absence of a universal normalization; \(R_D(a)\) supplies it for the answering setting.

10. How this was checked

The rate computations were verified numerically: a Blahut–Arimoto implementation reproduces the closed forms for binary and \(k\)-ary tasks to \(10^{-15}\), is invariant under the addition of score-irrelevant answer dimensions, and returns zero when a constant answer suffices.

The theorems were audited by four independent adversarial referees instructed to refute them. The audit rejected the first version of the main claim: the swap machine of Section 4 was constructed independently by two referees as a counterexample to a draft whose abstract omitted the copy-semantics condition, and the condition was promoted from a correction term to a necessity theorem. The finite-time audit re-derived the speed limit independently, tested the displacement lemma on \(2\times10^5\) random instances and the strengthened phase inequality on 60 randomized driven protocols with no violations, and closed one genuine gap in the lemma's proof. The version linked here is the one that survived.

11. Machine-checked in Lean

Numerical checks and adversarial referees reduce the probability of error; a proof assistant removes a category of it. The algebraic and information-theoretic skeleton of the main theorem is formalized in Lean 4 and machine-checked against Mathlib (both pinned to v4.30.0). The file compiles with no errors and no warnings.

The boundary between what the compiler proves and what it assumes is drawn honestly. The three per-phase second-law inequalities and the data-processing inequality enter as declared hypotheses, because Mathlib has no theory of Markov-jump entropy production; formalizing them would be a separate project in stochastic thermodynamics. Everything downstream of those hypotheses is verified by the compiler: the boundary identities, the telescoping cancellation where a hand sign-error would hide, the sign conditions, and, most pointedly, the converse \(R\le I(Q;\hat Y)\) reduced to csInf_le — the defining property of an infimum. That is the paper's claim that the converse needs no coding theorem, now a line the machine checks rather than a sentence a reader trusts.

So the formalization is conditional: given the second law in per-phase form and the standard data-processing inequality, the bookkeeping and the converse are correct with certainty. That is precisely the layer where a hand proof could have erred. Source: LandauerFrontier.lean, with notes on what is proved versus assumed.

Full statements, proofs, and the audit trail: draft v0.3 (PDF).