What thermodynamic computing actually proposes
Every computer sits in a bath of thermal noise. The design question is what to do about it, and this approach answers differently from everything before it.
Noise becomes the power source
Thermodynamic computing uses the same fluctuations that classical and quantum machines work to suppress, and treats them as what drives the computation. That inversion is the whole idea.
The framing matters for judging the energy claims. This is not a more efficient way to do the same operations; it is a different physical basis for computing, in which randomness is the resource rather than the error term. Its advantage should be expected on workloads that are already probabilistic, and not everywhere.
"But thermodynamic computing, a branch of unconventional computing, inverts the paradigms of both and uses those same fluctuations as its power source."
Classical chips pay a permanent tax to stay stable
A conventional device has to be held firmly in one state or the other, and holding it there costs energy continuously. That expenditure buys reliability, not computation.
Berkeley Lab illustrates the scale of ordinary computing's energy footprint with a single Google search consuming enough energy to power a six-watt LED for three minutes. Multiply that across an AI workload and the suppression tax is a large fraction of what a data center pays for.
"Modern computing requires energy: a single Google search, for example, consumes enough energy to power a six-watt LED for three minutes."
What is being built, and what is still open
This is where the English coverage and the research literature describe different states of the world.
Extropic has chips, a foundry plan and federal money
Extropic announced a letter of intent with the US Department of Commerce for up to $75 million through the CHIPS Research and Development Office. Its chips are described as using the thermal fluctuations of standard CMOS transistors to sample from programmable probability distributions.
Two details make this more than a funding headline. Using standard CMOS means no exotic fabrication is required, and the company describes a successor chip fabricated at a US foundry, which is a manufacturing commitment rather than a research milestone.
"TSUs use the natural thermodynamic fluctuations of standard CMOS transistors to sample directly from programmable probability distributions in a power-efficient manner, a capability useful for generative AI, simulations of biology, markets, and beyond."
The research position is that hardware is the open problem
The Berkeley Lab work states that working out how to realize these designs in hardware is what matters next. That is a description of unfinished business, published in the same year.
The two statements are compatible, and the reconciliation is worth being precise about: a company can be fabricating a chip built on one set of design choices while a research group holds that the general design-and-training problem is unsolved. What it means for a reader is that "does the hardware exist" has no single answer yet — it depends whose approach is being asked about.
"According to Whitelam, it's important to work out how to realize these designs in hardware."
The equilibrium constraint was the earlier blocker
Earlier thermodynamic computers had to settle into their lowest-energy configuration before an answer could be read, which caps how fast they can run. Removing that requirement is the advance the 2026 paper reports.
When the components themselves are nonlinear, the machine can be trained to produce nonlinear results at chosen times rather than at equilibrium. That converts a wait-for-settling device into one that can be clocked, which is the difference between a demonstration and something schedulable.
"We simulate a digital model of a thermodynamic neural network, and show that its parameters can be adjusted by genetic algorithm to perform nonlinear calculations at specified observation times."
What the trade-offs mean in practice
Two properties decide whether this technology fits a given workload, and neither is about raw speed.
Every run gives a different answer
A thermodynamic computer is stochastic: no two runs look the same, and the training methods used for digital neural networks do not apply. The research team used a genetic algorithm instead.
This is a real constraint rather than a detail. Workloads that need a reproducible answer are a poor fit; workloads that are sampling from a distribution anyway — generative models, simulation, probabilistic inference — are the natural target, which is the same territory the commercial claims point at.
"A thermodynamic computer is a stochastic system, meaning that no two runs on a thermodynamic computer look the same, and the methods used for training digital neural networks don't apply."
Expensive to train, cheap to run
The training framework costs considerably more than digital methods, and the resulting machine runs on very little energy. That asymmetry is the economic case.
It also tells you where the technology lands first. A device that is costly to prepare and cheap to operate suits a fixed function running constantly — inference at scale, an always-on sensor, a simulation loop — rather than work that gets retrained frequently. The economics reward stability, not flexibility.
"This training framework is considerably more costly than the methods used to train digital networks, but it yields a computer that can operate using very little energy after it's built"
The primary material here is split between a lab news page, a journal article and a company writeup, and each states its claims with different scope. Converting them to markdown before reading them side by side keeps each claim attached to its source, which is the only way to see that they are describing different things.
The honest position in mid-2026 is that thermodynamic computing has a working theory, a published training method and at least one company fabricating hardware, while the research community still treats general hardware realization as open. For anyone assessing it, the question worth asking is not whether the physics works — it does — but whether a workload can tolerate answers that differ every run. That constraint, not the energy figure, decides where this technology can go.
