Intelligence converges on sparsity across every substrate — biological neurons, artificial neural networks, neuromorphic chips — because entropy disposal is the binding constraint. The convergence is not analogy. It is physics. The AI industry is rediscovering what evolution solved four hundred million years ago.
A study published March 9 by researchers at UC Riverside, the Rochester Institute of Technology, and Caltech found that U.S. data centers could need 697 million to 1.45 billion gallons of new peak water capacity per day by 2030. The infrastructure cost: ten to fifty-eight billion dollars. The figure rivals the daily water supply of New York City.
The researchers' central finding was not the total volume but the ratio. Daily water demand from evaporative cooling systems spikes six to ten times above average usage, and for some planned facilities the multiplier exceeds thirty. Annual figures hide the constraint. The crisis is not that data centers use too much water on average. It is that on the hottest days of the year, they need infrastructure sized for a peak that may last hours — infrastructure that sits mostly idle for the remaining three hundred and sixty days.
This is not a story about water. The Grid tracked the power pledges. The Crucible tracked the supply shock. The Fiber tracked the photonics race. Those entries covered economics — who pays, who builds, who profits. This entry is about what makes the economics inevitable: the physics of heat.
The Tax
Every computation that destroys information dissipates energy. Rolf Landauer proved this in 1961. The lower bound is kBT ln 2 per erased bit — Boltzmann's constant times the ambient temperature times the natural logarithm of two. At room temperature, roughly 2.9 times ten to the negative twenty-one joules per bit. The number is vanishingly small. It is also absolute. No engineering can eliminate it. It is a consequence of the second law of thermodynamics.
Modern processors operate at roughly a million times the Landauer limit. The gap is enormous and has narrowed by orders of magnitude over five decades of chip design. But the direction is fixed. Every irreversible computation pays a tax in heat. The tax cannot be evaded. It can only be managed.
A data center is a building designed to manage this tax at industrial scale. The GPUs inside perform trillions of irreversible operations per second. Each operation pays. The accumulated heat must be removed or the chips throttle, degrade, and fail. The cooling infrastructure — chillers, cooling towers, liquid distribution units, heat exchangers — exists for one purpose: to export entropy from the building faster than computation generates it. The water is the vehicle.
This is why the constraint migration in AI infrastructure follows the sequence it does. Chips were the first bottleneck — TSMC's fabrication capacity limited how many GPUs could exist. Energy was the second — each rack now draws over a hundred kilowatts, up from ten a few years ago. Water is the third, because power becomes heat and heat must leave. The constraints are not independent. They are linked by thermodynamics: computation requires chips, chips require power, power generates heat, heat requires cooling, cooling requires water. The billions in delayed projects and canceled facilities are the financial expression of a thermodynamic chain.
What Twelve Watts Solved
The human brain arrived at this problem four hundred million years ago.
The brain consumes approximately twelve watts — twenty percent of the body's total energy budget for two percent of its mass. Blood flow per neuron is roughly invariant across mammalian species. Brain metabolic rate scales with volume at the five-sixths power — steeper than the body's three-quarter scaling. Per unit volume, the brain generates more heat than almost any other organ.
The constraint this creates is spatial. Blood vessels and neurons compete for physical space inside the skull. Every cubic millimeter devoted to cooling infrastructure is a cubic millimeter unavailable for computation. The brain cannot simply fire more neurons to think harder — the cooling system would need to grow, displacing the neurons it was meant to support. There is a ceiling imposed not by neural capacity but by the geometry of the skull and the fluid dynamics of blood.
Evolution's solution was sparsity. At any given moment, the vast majority of cortical neurons are silent. The neural efficiency hypothesis — confirmed across two decades of neuroimaging research — shows that higher intelligence correlates with less brain activation, not more. The most efficient brains do not process more. They process less, and more selectively. Sparse coding maximizes the ratio of cognitive work to metabolic cost by activating only the neurons whose contribution exceeds the thermodynamic cost of their activation.
This is not a design choice in the way an engineer makes design choices. It is the only architecture that works under the constraint. A dense-firing brain would generate heat faster than blood could remove it. Seizures — episodes of synchronized dense neural firing — demonstrate what happens when the sparsity constraint is violated. The system destabilizes within seconds.
What Six Hundred Billion Dollars Is Rediscovering
Every frontier AI model now uses a sparse architecture. Claude, GPT, Gemini, DeepSeek, Llama, Mixtral — all employ Mixture of Experts, activating one or two expert subnetworks per token out of sixty-four to over a thousand available. DeepSeek V3 activates thirty-seven billion parameters per token from a total of 671 billion — a sparsity ratio of roughly five percent. Google's GLaM demonstrated that a sparse model with 1.2 trillion total parameters and sixty-four experts outperformed a dense 175-billion-parameter model on zero-shot benchmarks while using half the inference compute.
The industry arrived at sparsity through economics. Dense models cost too much to run. Activating every parameter for every token requires proportional energy, proportional cooling, proportional water. Mixture of Experts reduces inference cost by activating only the relevant subset. The economic pressure is a proxy for the thermodynamic pressure: less compute per token means less heat per token means less water per token.
The convergence is striking. The brain activates a small fraction of its cortical neurons per cognitive operation. Frontier AI models activate roughly three to five percent of total parameters per token. Both systems arrived at the same architectural solution independently — one through four hundred million years of natural selection under metabolic constraint, the other through five years of economic pressure under infrastructure constraint — because both face the same underlying physics. Entropy disposal is the bottleneck. Sparsity is the response.
The Work Ratio introduced Peter Fagan's central metric: chi equals goal-directed work divided by irreversible information processed. That entry asked why most AI deployments produce zero return. The Heat Tax is about what the same framework predicts for architecture. Fagan's Conservation-Congruent Encoding generalizes Landauer's principle to arbitrary conserved quantities via metriplectic flows — the cost of erasing information extends beyond thermal channels to any conserved quantity whose conjugate variable enforces the erasure cost. The framework is substrate-neutral. It applies to neurons, transistors, and any future computing substrate, because it derives from conservation laws rather than engineering choices.
The prediction follows from the physics. Intelligence that persists does so by maximizing the numerator — goal-directed work — while minimizing the denominator — irreversible processing. Sparsity minimizes the denominator. The convergence of biological and artificial intelligence on sparse architectures is the convergence on the same physical optimum: the architecture that extracts the most work per unit of entropy produced.
The Third Path
Neuromorphic computing confirms the convergence from a third direction.
Intel's Loihi 2 runs a sparsified deep learning model at one hundred and fiftieth the energy of a GPU on the same video processing task. UC San Diego's brain-inspired chip — which co-locates memory and computation on the same substrate, eliminating the von Neumann bottleneck — achieves a thousandfold power efficiency improvement for real-time sensory processing. IBM's NorthPole demonstrates twenty-five times the energy efficiency of an H100 GPU for image recognition. Intel's Loihi 3, fabricated on a four-nanometer process and announced in 2026, operates at a peak load of 1.2 watts for tasks that would require hundreds of watts on a GPU.
These are not incremental improvements. They are architectural convergences on the same solution evolution found: sparse activation, local computation, physical proximity of memory and processing. The brain has no von Neumann bottleneck because neurons store and process in the same structure. Neuromorphic chips reproduce this layout. The efficiency gains come not from better engineering of the existing architecture but from switching to the architecture that thermodynamics favors.
Three substrates — biological neurons, digital transistors, neuromorphic hardware — each designed under different constraints, by different processes, at different timescales. All three converging on the same architectural features: sparse activation, co-located memory and compute, and the lowest entropy production per unit of useful work the physics allows. The convergence is the evidence that the constraint is real and the solution is not optional.
What Comes After Water
The constraint migration is accelerating because each successive bottleneck has a higher peak-to-average demand ratio. Chips have roughly constant demand — a data center needs the same GPUs in January and July. Energy demand varies modestly by season. Water demand spikes six to ten times above average on the hottest days, and some facilities see a multiplier above thirty. Each step up the thermodynamic chain encounters more temporal variability, which means more expensive infrastructure per unit of average capacity.
Sightline Climate reports that thirty to fifty percent of large data centers scheduled for 2026 face delays. At least nine projects have been fully canceled. Twenty-five percent of the hundred and forty tracked projects have not disclosed their power sourcing plans. The physical world is asserting that entropy disposal has a geography, a hydrology, and a politics.
The trajectory points in a specific direction. As cooling constraints tighten, AI sparsity ratios will increase — fewer active parameters per token, less heat per inference, less water per query. Current frontier models activate roughly three to five percent of their total parameters. The thermodynamic pressure pushes that number down. The economic pressure pushes in the same direction. Both converge on sparser, cooler, more efficient.
The brain found its floor over hundreds of millions of years of selection pressure. The question is how quickly artificial intelligence descends toward it — and what the industry looks like when it arrives. The six hundred and fifty billion dollars in AI infrastructure spending this year is, in thermodynamic terms, an investment in entropy export capacity. The return on that investment depends not on how much computation the infrastructure enables but on how little entropy each computation produces.
The heat tax is the line item nobody put in the budget. It was always there. Landauer wrote the invoice in 1961. The bill is arriving now.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)