Extropic's Guillaume Verdon on thermodynamic chips, a US government LOI, and why GPUs aren't the endgame

Extropic's thermodynamic bet

Guillaume Verdon's core argument is that GPUs are not the endgame. His company, Extropic, is building a new class of chips designed to run probabilistic AI workloads — diffusion models, predictive control, signal processing — by exploiting the physics of electronics rather than fighting it.

The logic is straightforward enough to stress-test. Conventional chips burn enormous power maintaining deterministic state. Generative AI workloads are inherently probabilistic. Verdon argues it is wasteful to run a probabilistic algorithm on hardware engineered to never be uncertain. Extropic's chips operate probabilistically at low power and map algorithms like diffusion models directly onto the hardware's physical behavior.

“At XTropic, we've been pioneering this new form of computing from the ground up called thermodynamic computing... We reinvented how to leverage the transistor for the era of generative AI. We demonstrate that in principle we can manufacture these chips in the US. And that was very interesting to all sorts of folks higher up... This announcement is with the chips R&D office.”

US fab angle

The geopolitical pitch may be the most immediately actionable part of the story. Because thermodynamic chips don't require the smallest transistors — probabilistic operation means far fewer transistors are needed overall — they can be manufactured on older-node fabs in the US rather than cutting-edge offshore facilities. Extropic has taped out with both TSMC and GlobalFoundries to demonstrate this. The CHIPS R&D office signed a letter of intent on the back of it, with the government taking an equity stake rather than offering a pure grant — Verdon frames this as taxpayers getting upside, not just writing checks.

A product announcement around "Z1" and associated stacks is expected next week.

Sparse transformers

Verdon teases a forthcoming blog post on a sparsity scaling law. The claim: Extropic's chips handle sparse matrix operations efficiently, meaning a sparsified transformer with the same parameter count as a dense model can reach comparable performance, but at far lower watts per FLOP. The practical implication is that even conventional transformer workloads could run on Extropic's architecture if models are structured to match it.

RTX Ventures is listed as an investor, alongside other defense primes, with interest concentrated in physical intelligence, predictive control, and signal-processing applications — workloads Verdon says would typically go to FPGAs.

Cycle timing

Verdon is explicit that Extropic is starting the next S-curve while the GPU cycle is still running. His read is that going all-in on the current paradigm without hedging into the next generation of compute substrates is how companies get wiped out when the cycle turns. He uses current AI tooling to accelerate Extropic's own development — his framing is "new game plus," speedrunning fifteen years of deep learning progress with AI as a power-up.

The near-term credibility question is the gap between thermodynamic computing's theoretical efficiency gains and production-scale deployment. The LOI and the GlobalFoundries tape-out are the most concrete proof points on the table. The Z1 announcement next week will be the next signal worth watching.