Every conventional computer chip fights a constant war against a fundamental nuisance: the random jittering of atoms and electrons caused by thermal energy. This noise limits how small transistors can be scaled and forces chips to consume vast amounts of energy just to maintain reliable operation.
A small but growing community of physicists and engineers is asking whether this war is worth fighting, or whether noise could be turned into a computational resource instead. The result is a nascent field called thermodynamic computing, which harnesses random thermal fluctuations rather than suppressing them.
As Patrick Coles, chief scientist at Normal Computing, put it: “The field is about designing computers that exploit thermodynamics as a computational resource.”
How it works
The core insight comes from physics. A simple electrical circuit, a network of resistors, capacitors, and inductors, will naturally explore many possible states as its components are buffeted by thermal noise. If the circuit is designed so that the most stable state corresponds to the solution of a mathematical problem, then the system will find that solution simply by settling into equilibrium.
This is analogous to protein folding, where a long chain of amino acids “wiggles” randomly until it finds the shape with the lowest energy. Nature has been doing this kind of computation for billions of years.
Stephen Whitelam, a staff scientist at Lawrence Berkeley National Laboratory, demonstrated the principle by building a circuit that could denoise images, reconstructing a recognizable image from random static, while dissipating roughly 100 billion times less heat than a conventional digital neural network doing the same task. “The computer designs we’ve come up with so far for thermodynamic computing are only as capable as the small digital neural networks of around 1990,” Whitelam told Quanta Magazine.
Two approaches
Two distinct strategies have emerged. Normal Computing, a startup founded by former Google Brain and Google X researchers, builds equilibrium systems that use networks of RLC circuits (resistors, inductors, capacitors) to perform calculations through their natural thermal fluctuations. The company published a peer-reviewed demonstration in Nature Communications in 2025 and has since taped out its first dedicated thermodynamic chip, the CN101, designed for running multi-modal diffusion models.
Extropic, a second startup founded by ex-Google quantum researchers, takes a nonequilibrium approach. Its XTR-0 platform uses thousands of semiconductor components driven by injected noise to perform probabilistic calculations. The company claims the approach could use 10,000 times less energy than conventional hardware for generative AI inference, though the results have been published only as a preprint and have not yet been peer-reviewed.
Where it might go
Thermodynamic computing is not a replacement for digital computers. It is best suited for a specific class of problems: those involving probability, sampling, and optimization, precisely the kind of calculations that dominate modern AI. Diffusion models, Bayesian inference, and molecular simulations all involve exploring large spaces of possibilities, and all could benefit from hardware that does this naturally rather than through brute-force digital simulation.
The field is early. Normal Computing’s roadmap calls for increasingly capable chips over the next few years, but even its own researchers compare the current state to quantum computing in the 1990s. Whitelam offered a perspective from biology: “To my physicist’s way of thinking, it would be fair to say that nature uses Langevin computers programmed by evolution.”
The question now is whether researchers can build engineered versions of those biological computers, and make them useful.
Sources
- Quanta Magazine: “Thermodynamic computers go with the energy flow” (July 15, 2026)
- Melanson, D., et al. “Thermodynamic computing via equilibrium fluctuations.” Nature Communications (2025). DOI: 10.1038/s41467-025-58199-y
- Whitelam, S. “Generative learning of nonequilibrium thermodynamic computers.” Physical Review Letters (2026)