A physicist at Lawrence Berkeley National Laboratory just demonstrated something that sounds impossible: a computer that generates images the same way AI does, but uses the heat it would normally waste as its primary fuel. The theoretical energy savings are eleven orders of magnitude. Ten billion times less power than a GPU running the same task.
Stephen Whitelam published the results in Physical Review Letters on January 20. A companion paper with co-author Casert landed in Nature Communications ten days earlier. Together, they describe a new class of machine — a generative thermodynamic computer — that produces structured images from random noise without a neural network, without backpropagation, and without the electricity bill that comes with both.
The proof of concept generates handwritten digits. MNIST, the dataset every machine learning student trains on first. That's modest. But the mechanism underneath is not.
How It Works
Modern AI image generators — DALL-E, Midjourney, Stable Diffusion — are diffusion models. They learn to reverse noise. You take a photo, add static until it's unrecognizable, then train a neural network to undo each step. At generation time, you feed the model pure static and it hallucinates an image into existence, one denoising step at a time. Every step requires matrix multiplication across billions of parameters. Every matrix multiplication burns watts.
Whitelam's machine skips the neural network entirely. Instead, it encodes the denoising instructions into the physical dynamics of a thermodynamic system — electrical circuits whose components naturally fluctuate due to thermal noise. The same randomness that engineers spend billions trying to eliminate from chips becomes the computational engine.
Training works by maximizing the probability that the system can reverse a noising trajectory. The computer learns to undo decay not through gradient descent but through the physics of its own hardware. When it generates an image, the thermal fluctuations do the work that billions of floating-point operations would otherwise perform.
The energy comparison is where the numbers get absurd. A single image generation on a GPU requires trillions of operations, each burning energy. Whitelam's thermodynamic system performs the equivalent computation through natural physical processes that dissipate minimal heat. The theoretical gap: ten billion to one.
The Catch
"We don't yet know how to design a thermodynamic computer that would be as good at image generation as, say, DALL-E," Whitelam told IEEE Spectrum. He's generating 28-by-28-pixel handwritten digits. DALL-E generates photorealistic scenes at 1024-by-1024. The complexity gap between those two tasks is enormous.
And "theoretical" is doing heavy lifting in that ten-billion figure. "Near-term designs will be something in between that ideal and current digital power levels," Whitelam said. The physics permits eleven orders of magnitude improvement. The engineering might deliver three or four. Even that would be revolutionary.
The bigger problem is scaling. Whitelam's system uses coupled degrees of freedom — physical components whose interactions encode learned patterns. Scaling to millions of parameters (let alone billions) requires fabricating analog circuits of extraordinary precision. Normal Computing, a New York startup, has built a prototype chip with eight resonators. Eight. GPT-4 has roughly 1.8 trillion parameters.
Why It Matters Now
Extropic, another startup in the thermodynamic computing space, claims their Thermodynamic Sampling Units can run denoising models at 10,000 times lower energy per operation than GPUs. Their Z1 chip, built on standard CMOS transistors, is expected in early access this year. Unlike exotic academic approaches requiring magnetic junctions or optical systems, Extropic uses the natural thermal noise of ordinary transistors. The entire chip is fabbed at existing semiconductor plants.
The timing is not accidental. AI's power consumption has become a political crisis. PJM Interconnection, which manages the grid for 65 million Americans, fell 6,625 megawatts short of its reliability target last year — the first time the entire regional grid missed its benchmark. Data centers accounted for 97 percent of the new demand. Fourteen states have data center moratorium movements. Bernie Sanders and Ron DeSantis have both introduced legislation targeting AI's electricity footprint.
Gartner projects $2.5 trillion in global AI spending this year, up 44 percent. The industry's growth trajectory assumes cheap, abundant power. That assumption is collapsing.
If thermodynamic computing delivers even a fraction of its theoretical promise — not ten billion times, but a thousand times, or even a hundred — the economics of AI infrastructure change completely. A hundred-fold reduction in power means a data center that consumes 100 megawatts could run on one. It means the grid crisis stabilizes. It means AI stops being a geopolitical liability measured in gigawatts.
That's a big if. The field is early, the prototypes are primitive, and the gap between handwritten digits and frontier AI models is measured in decades of engineering. But the physics is real, the math checks out, and the industry desperately needs it to work.
Nobody in AI talks about thermodynamic computing yet. Within five years, they won't talk about anything else.
Sources: Physical Review Letters (Vol. 136, 037101), Nature Communications (17, 1189), IEEE Spectrum, Tom's Hardware, Live Science, Lawrence Berkeley National Laboratory
Top comments (0)