Extropic

building the ultimate substrate for intelligence

Extropic designs probabilistic computing hardware for generative AI, using arrays of pbits fabricated in standard CMOS to sample from distributions natively in silicon rather than emulating that behavior on deterministic GPUs. The company is led by founder and CEO Guillaume Verdon, previously of Google Quantum AI , alongside CTO Trevor McCourt, who came from Alphabet's X moonshot lab and AWS . This launch matters because the team is finally moving past the X0 research chip and XTR0 dev kit into a full stack pitched at working AI developers. The centerpiece is Z1, described in the video as a taped-out chip with roughly 269,000 pbits arranged in a sparse graph, drawing under one watt in a package smaller than 12mm on a side. Z1 will ship in two form factors, an M.2 "thermocompute stick" with over half a million pbits aimed at edge, AR, and robotics workloads, and a PCIe accelerator card packing more than four million pbits for workstations and servers running alongside conventional accelerators. Early access is planned for 2027, with a first pbit cluster targeted for the following year. Alongside the hardware roadmap, Extropic is releasing Torx, an open source framework for stochastic differentiable programming that lets developers build directed graphs of stochastic transition kernels and train them end to end with gradient descent, positioned as the PyTorch-style layer above their lower level Thermal library. A companion white paper on Thermalizers describes how Torx programs get compiled into energy based models that map onto thermodynamic devices, and an open access API lets developers run Torx and Thermal programs today on GPU simulators while waiting for Z1 silicon. For the workloads Extropic targets, the team claims up to 10,000x greater energy efficiency than GPUs, a number worth testing rather than taking at face value, but one that explains why the company is worth watching as AI power budgets keep climbing.

Tags

Comments (15)