Extropic has signed a letter of intent with the US Department of Commerce for funding of up to $75 million through the CHIPS Research and Development Office. The goal is to develop and manufacture its thermodynamic AI chips on American soil, scaling a technology from prototype to industrial production — one that harnesses the natural thermal fluctuations of CMOS transistors to perform probabilistic computation.
At the heart of the project are Thermodynamic Sampling Units (TSUs), processors designed to directly sample programmable probability distributions by exploiting the intrinsic thermal noise of standard transistors. This seemingly counterintuitive approach enables workloads typical of generative AI, biological simulation, and financial markets to run at a fraction of the energy consumed by conventional GPUs. On several key generative AI benchmarks, efficiency gains reach orders of magnitude. The plan also includes mass production of the Z1 processor, the development of rack-format thermodynamic systems, and the construction of an American manufacturing supply chain for the next-generation Z1.5 chip, built on mature process nodes.
The backdrop is an AI race colliding with the physical limits of the power grid. The growth of generative models demands ever-increasing power, and watt availability has become a genuine bottleneck for the US computing industry. Extropic explicitly frames its work within America’s broader energy strategy: as the country moves to secure energy supply through a nuclear renaissance — driven by presidential executive orders issued in May 2025 — chip efficiency will determine how much value can be extracted from every watt produced. Guillaume Verdon, Extropic’s founder and CEO, captured the issue plainly: energy is the defining constraint of the current AI era, and pushing the deterministic digital paradigm to its limits cannot be the ultimate answer.
The agreement is part of a broader Department of Commerce initiative that announced letters of intent with seven companies totaling $874 million for research and development across the advanced computing semiconductor supply chain. The underlying logic mirrors the public-private partnerships that built the American semiconductor industry: government as both first customer and first investor, capable of bridging the gap between a working prototype and full-scale production. Extropic fits squarely within this tradition, bringing an architecture radically different from GPUs — one that abandons deterministic digital logic in favor of stochastic electronics.
If TSUs in production confirm the advantages already observed at the prototype stage, the impact will extend well beyond data centers. The probabilistic architecture lends itself to robotic systems, autonomous platforms, and military sensors that must reason in real time under tight energy constraints. Nuclear power provides the watts; thermodynamic chips multiply what those watts can achieve. On paper, at least, the chain holds.
