Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-20 · high
It uses p-bits to run the sampling and optimization work GPUs own, but it's a research machine, not a product yet.
Why now: It surfaces in the July 20 coverage as one of the only non-software answers yet floated for AI's energy cost problem.
A new probabilistic computer treats thermal noise as a computational resource instead of an error to correct, and IEEE Spectrum calls it the largest machine of its kind built so far. Instead of fighting the randomness silicon generates at room temperature, it runs on p-bits, probabilistic bits that use that noise directly to solve sampling and optimization problems.
That's the same class of work GPUs own. Most inference and optimization jobs burn GPU power to do this kind of math, and energy cost is turning into the industry's real constraint on scaling AI. A different physical substrate that handles the same problems is one of the few answers to that cost that isn't just a software optimization.
Research hardware, not a product yet. Spectrum's coverage doesn't say how it compares to a GPU on cost or power per operation, and it doesn't say when, or if, this leaves the lab. I'd want that number before believing the energy story holds up outside a controlled setup.
My bet: the fix for AI's energy problem won't come from a faster GPU or a smarter kernel. It'll come from someone swapping out the substrate entirely, the way this does. Everything else is optimization on top of the same physics.
Worth watching for the first benchmark against a GPU doing the identical sampling job. That number doesn't exist yet.
Each link below shares sources, entities, or timing with this story.
Shared entity: GPUs / Shared topic / Earlier coverage / Tension
Both cover GPUs; overlapping topics (cost, inference); earlier GPUs coverage from 2026-06-02.
Shared entity: GPUs / Shared topic / Earlier coverage / Downstream implication
Both cover GPUs; overlapping topics (gpus, inference); earlier GPUs coverage from 2026-04-10.
Shared entity: GPUs / Shared topic / Earlier coverage
Both cover GPUs; overlapping topics (gpus, hardware); earlier GPUs coverage from 2026-06-29.
Both cover GPUs; overlapping topics (above, gpus); earlier GPUs coverage from 2026-06-08.
Both cover GPUs; overlapping topics (cost, inference); earlier GPUs coverage from 2026-06-07.
Shared entity: IEEE Spectrum / Same source domain / Earlier coverage
Both cover IEEE Spectrum; reported by the same outlet (spectrum.ieee.org); earlier IEEE Spectrum coverage from 2026-05-19.
Shared entity: GPUs / Shared topic / Earlier coverage
Both cover GPUs; overlapping topics (gpus, largest); earlier GPUs coverage from 2026-05-11.
Both cover GPUs; overlapping topics (gpus, hardware); earlier GPUs coverage from 2026-04-21.