Extropic's thermodynamic-chip bet lands a $75M CHIPS R&D letter of intent

Builds Thermodynamic Sampling Units (TSUs): probabilistic AI chips that sample from energy-based models directly, trading transistors for physics to cut power.

That computing with physical noise — thermodynamic sampling, not digital matrix multiplication — beats GPUs on energy for generative AI, on mature nodes.

$14.1M seed (Dec 2023), plus up to $75M planned CHIPS R&D funding under the non-binding July 2026 LOI.

Extropic was founded in 2022 on the bet that energy, not compute, would become AI's binding constraint — a claim its founder repeated when the power scramble around AI data centers made it conventional wisdom.

It surfaced in Dec 2023 with a $14.1M seed; a March 2024 manifesto drew a 185-point HN debate. On Oct 29, 2025 it unveiled the X0 prototype silicon (two variants taped out and tested at two different fabs), the XTR-0 development platform, the Denoising Thermodynamic Model algorithm, and the thrml simulator library — then spent 2026 turning that into a federal funding case.

AI's binding constraint has moved from FLOPs to watts; a chip that buys intelligence per joule attacks the bottleneck every new data center feels.

Skipping leading-edge nodes turns a hardware bet into an industrial-policy match — idle mature-node US capacity is exactly what CHIPS R&D exists to fund.

Sampling, not matrix multiplication, is the workload TSUs own: a narrow beachhead instead of a head-on GPU replacement.

Physics-native randomness does for free what digital circuits burn energy to fake — noise is the feature, not the bug.

A deep-tech wedge can be regulatory as much as technical: building on mature nodes made Extropic fundable as onshoring, not just faster AI. Narrow your workload claim until the hardware can prove it.

As of September 2026, Extropic has prototype silicon proven at two fabs, a developer platform in partner hands, and a non-binding LOI for up to $75M in CHIPS R&D funding signed July 30, 2026 — aimed at the Z1 production chip, rack-scale clusters, and US-foundry fabrication of Z1.5. No definitive award has been announced, revenue-scale customers are unnamed, and the orders-of-magnitude efficiency claims remain self-reported pending third-party benchmarks on Z1 hardware.