A physics yardstick for computing that prices brain and silicon alike

What the work claims

The award's premise is that every computing device, from a smartphone to a data center to the human brain, pays an energetic cost to process information, and that for a system of connected computers the minimum cost depends crucially on how those computers are connected.¹ The central analytical tool is Mismatch Cost, abbreviated MMC, which the record defines as a strictly positive lower bound on thermodynamic cost, derived under the investigator's prior NSF-funded work.¹ The load-bearing property is that MMC is hardware-agnostic: it applies across physical substrates, from conventional semiconductor circuits to neuromorphic hardware to biological neural tissue, and therefore, the record argues, results about MMC are guaranteed to remain relevant no matter how technology evolves.¹

What is established, per the award abstract, is that MMC depends crucially on network topology; what is not known is the shape of that dependence. The project's program is to pin it down. Because the topology-computation relationship must be studied under precise control of both variables, the work uses trained neural networks and restricted Boltzmann machines as stylized models of distributed computers, trains them on well-defined tasks including instances of the canonical KSAT benchmark from computer science, systematically varies the internal connectivity, and then characterizes how thermodynamic cost, computational speed, and robustness to component failure jointly depend on network structure.¹ The framing is explicit about ambition: the results are meant to provide the scientific foundation for far more energy-efficient computers and AI hardware, and possibly fundamental insight into how the human brain operates at such small energy cost.¹

Read this as what it is: a theoretical physics award, active from October 2026, with no project results yet. The novelty claim is not MMC itself but the systematic mapping of MMC against topology in controlled computing systems, and the wager that the resulting theory will transfer across substrates.

How it works

The physics lives in nonequilibrium thermodynamics. Any real computation is a physical process, and physical processes that run away from thermodynamic equilibrium dissipate energy. Mismatch Cost is a bound on part of that dissipation: roughly speaking, it quantifies the excess cost a system pays when its internal dynamics are mismatched to the computation it is actually performing. Because the bound is derived from statistical mechanics rather than from any model of a transistor or a neuron, it carries no assumptions about what the system is built from. That substrate-neutrality is the entire point, and it is what lets the award treat a trained neural network, a restricted Boltzmann machine, and, in principle, a living neural network as instances of one mathematical object: a network of computing components whose wiring shapes its minimum energy bill.

The experimental strategy is equally deliberate. Biological brains cannot be rewired on demand, and chip manufacturers do not publish thermodynamic audits, so the project builds its testbeds in silico: artificial networks whose connectivity the researchers control completely, trained on benchmark tasks such as KSAT, a canonical satisfiability problem used to study hard computation. Vary the wiring, measure the bound, and look for the topology-cost-speed-robustness trade space. If regularities emerge, they are candidates for universal laws of computing cost, and the brain can then be examined as one more data point rather than as an inaccessible special case.

Where a skeptic should push

The single most load-bearing assumption is that a lower bound derived in statistical mechanics tells you something decision-relevant about real computers. A lower bound is a floor, not a bill: it says what a system must dissipate at minimum, not what it does dissipate, and the gap between floor and reality is where all the engineering lives. Silicon systems are within a few orders of magnitude of their physical limits in some domains and far from them in others; for biological neural tissue nobody has a trustworthy measurement of actual dissipation per computation, partly because nobody agrees on what counts as one computation in a brain. A bound that cannot yet be compared to a measured quantity risks being precise and unfalsifiable at once.

Second, the testbeds are stylized. Restricted Boltzmann machines and trained neural networks on KSAT are well-controlled, but their topology is engineered and static, while cortical wiring is developmental, plastic, and metabolically coupled. If the topology-MMC relationship in artificial networks is itself complicated, transferring it to living tissue will require assumptions the award does not enumerate. Third, the speed and robustness axes introduce their own trap: a network topology that minimizes dissipation may do so by being slow or fragile, so any governance-flavored metric distilled from this work will be a multi-dimensional trade-off, not a single number, no matter how convenient a single number would be. Finally, the award's own rhetorical move, guaranteed to apply no matter how technology evolves, is exactly the kind of claim that ages badly; the guarantee is mathematical, but its relevance is empirical.

A substrate-neutral meter reaches neural tissue

For anyone tracking computation on living neural tissue, the opportunity is a benchmark that hype cannot bargain with. The energy-efficiency claim is the most durable argument for organoid and neural computing: brains compute on roughly twenty watts, orders of magnitude below silicon for some workloads. Today that claim is marketed, not metered; vendors compare cherry-picked workloads against cherry-picked baselines. A validated, hardware-agnostic lower-bound framework, extended from artificial networks to biological ones, would let a platform customer ask a vendor a much harder question: what does your substrate dissipate per unit of verified computation, relative to its physical floor, and how does that change with the wiring? Efficiency becomes a spec sheet quantity rather than a narrative. That would be genuinely useful to the field's honest players, because it punishes hand-waving.

The non-obvious implication runs the other direction, and it deserves more attention than it will get: a meter that is substrate-neutral is also a detector that is substrate-neutral. If physics can assign a rigorous thermodynamic cost to a computation regardless of whether it runs on silicon or on living tissue, then physics is, in principle, measuring how much computation a given chunk of matter is doing, full stop. Some accounts of moral status for organized biological systems lean, in part, on the quantity or sophistication of computation they perform. A governance framework that ever took that route would need a way to count computation that does not presuppose an answer to the neural-versus-silicon question; thermodynamic cost bounds are one of the very few candidate instruments with that property. Nobody should welcome that shortcut. Reducing a moral-status judgment to a single physical metric would repeat, in a new register, every history of governance-by-number, where the metric's apparent objectivity conceals the choices baked into its definition: per what operation, averaged over what time window, against which reference topology.

The threat model has three parts, all grounded in the mechanism. First, metric capture: if a thermodynamic efficiency index becomes the yardstick, whoever validates the measurements on living tissue writes the number that regulators cite, and validation is a vendor-shaped activity. Second, gaming by topology: since the bound depends on wiring, architecture choices can be tuned to the metric rather than to the customer, and a substrate optimized to look cheap on MMC-relevant axes may be worse on the axes the metric ignores, including the ones that matter for tissue welfare. Third, the obsolescence angle cuts against complacency in both directions: if silicon neuromorphic hardware keeps closing the efficiency gap, living-tissue computing must justify itself on measured cost rather than poetic analogy; and if living tissue does hold a real thermodynamic advantage, the same meter that proves it will also quantify exactly how much of that advantage a given experiment or platform is wasting. A yardstick this general does not care which substrate wins. It only makes the scorekeeping harder to fake, which is precisely why its eventual misuse needs to be designed against now, while it is still a theory project.

The bottom line

Established: Mismatch Cost is a real, published-in-spirit framework from the investigator's prior work, and the dependence of such thermodynamic bounds on network topology is a genuine open problem with a sensible research program attached. Hypothesis, not result: that the topology-cost-speed-robustness landscape mapped in artificial networks will transfer to biological neural tissue in a way that yields measurable, decision-relevant comparisons. What would confirm it: independent measurements of dissipation in a biological neural preparation that the framework predicts correctly, within stated definitions of a computation. What would break it: persistent orders-of-magnitude gaps between bound and measurement, or topology-cost relationships so task-specific that no transferable law exists. The governance takeaway is independent of the science: the moment a substrate-neutral computation meter exists, it will be pulled toward moral-status reasoning and vendor benchmarking alike. The field should want the meter and dread its first naive user, and it has about three years, the length of this award, to prepare for both.