Thermodynamics of Computation

In the AI Story

The second law of thermodynamics, formulated in the mid-nineteenth century through the work of Carnot, Clausius, and Kelvin, governs every energy conversion in the universe. No engine, biological or mechanical, converts input energy to useful work with perfect efficiency—some fraction always dissipates as heat. Computational devices are thermodynamic engines: electricity flows through transistors, charges capacitors, and performs logical operations, generating heat at every stage. Modern GPUs, optimized for the parallel matrix operations AI requires, draw 300-700 watts under load—comparable to household appliances. A data center housing tens of thousands of GPUs is, thermodynamically, a facility converting megawatts of electricity into computation and heat at rates that require industrial-scale cooling infrastructure.

Cooling system design trades off energy efficiency, water consumption, capital cost, and operational complexity. Evaporative cooling achieves the highest thermal efficiency by exploiting water's latent heat of vaporization—the energy required to convert liquid to vapor. The method is thermodynamically elegant and water-intensive: a single large data center can evaporate one to five million gallons daily. Air cooling eliminates water consumption but requires more electrical energy for fans and heat exchangers, particularly in warm climates where ambient temperature approaches server exhaust temperature. Liquid immersion cooling submerges servers in dielectric fluid, improving heat transfer and reducing parasitic energy losses, but requires hardware redesign and introduces maintenance challenges. Every configuration faces the same thermodynamic constraint: the heat must go somewhere, and moving it costs energy.

The Landauer limit, derived by Rolf Landauer at IBM in 1961, specifies the minimum thermodynamic cost of erasing one bit of information: approximately 2.9 × 10⁻²¹ joules at room temperature. Modern transistors operate roughly a million times above this limit, indicating vast potential for efficiency improvements. But the limit is non-zero—even a perfectly efficient computer operating at the Landauer bound would generate heat proportional to the number of bit-erasure operations performed. AI inference involves billions of such operations per query. The theoretical minimum energy cost is already significant at scale; practical energy costs are orders of magnitude higher. Efficiency improvements are real and necessary, but they move the floor, not eliminate it.

The thermodynamic constraint interacts with the Jevons Paradox to produce a compounding effect: as chips become more energy-efficient per operation, the cost of computation falls, demand expands, and total computation grows faster than per-operation efficiency improves. Smil documents this pattern across lighting (LED efficiency increased twentyfold, total lighting electricity tripled), refrigeration (unit efficiency improved 75%, total consumption held constant because units grew larger), automotive transport (fuel efficiency improved 30%, total gasoline consumption rose because people drove more). The pattern is structural, not incidental—efficiency that reduces cost tends to expand demand unless institutional constraints (price floors, quotas, regulatory caps) prevent the expansion. AI computation currently faces no such constraints, and demand shows no signs of saturation.

Origin

The thermodynamic analysis of computation traces to Landauer's 1961 paper "Irreversibility and Heat Generation in the Computing Process," which established that information erasure has a minimum energy cost. The framework gained practical significance with the rise of high-density integrated circuits in the 1970s-1980s, when chip designers confronted heat dissipation as a binding constraint on transistor density. The contemporary relevance emerges from AI workloads' extreme computational intensity: training a large language model involves 10²⁴-10²⁵ floating-point operations, each contributing to aggregate heat generation that data centers must manage at industrial scale.

Smil's specific application of thermodynamics to AI appears in his 2025 Pictet essay warning about exaggerated AI claims and his February 2026 Bankinter webinar quantifying the energy demands. The framework synthesizes his decades of work on energy systems, power density, and infrastructure constraints. His method—count the joules, follow the supply chain, measure the construction timelines—reveals that AI's "weightless" cognitive outputs rest on one of the heaviest industrial substrates in modern civilization, comparable in energy intensity to aluminum smelting or steelmaking when measured per unit of economic value created.

Key Ideas

Second law as binding constraint. Every computation converts electrical energy to heat according to thermodynamic law; efficiency improvements reduce but cannot eliminate this conversion, making heat management a permanent rather than transitional challenge.

Cooling-computation coupling. Data centers consume 30-40% of total electricity for cooling systems that remove computational heat; this overhead is not eliminable, only reducible through better chip design and cooling technology.

Landauer limit as theoretical floor. Even perfectly efficient computation has non-zero energy cost for bit erasure; current chips operate a million times above this limit, indicating efficiency headroom that is substantial but finite.

Heat scales with total computation. Aggregate heat generation from AI workloads grows when demand increases faster than per-operation efficiency improves—the thermodynamic expression of the Jevons Paradox applied to intelligence.

Water-energy-carbon triangle. Removing computational heat requires water (evaporative cooling) or extra electricity (air cooling); the electricity itself has embedded water costs (thermoelectric generation) and carbon costs (grid energy mix)—creating a three-way resource constraint.