Extropic is building thermodynamic computing hardware (extropic.ai)

Extropic unveiled thermodynamic computing hardware centered on thermodynamic sampling units (TSUs), which are inherently probabilistic processors designed to perform sampling and stochastic inference far more energy-efficiently than conventional GPUs. Their prototype platform, XTR-0, provides low-latency communication between Extropic chips and a traditional CPU, enabling hybrid workflows that offload sampling-heavy tasks to TSUs while keeping control and deterministic computation on the host. Alongside the hardware, Extropic released an open-source Python library for developing thermodynamic algorithms and simulating execution on TSUs, and is hiring engineers and scientists to build out the stack. For the AI/ML community this matters because many probabilistic workloads—Bayesian inference, MCMC, Boltzmann machines, stochastic optimization and sampling for generative models—are naturally matched to inherently stochastic hardware. If TSUs deliver on their energy claims, they could drastically reduce the cost and carbon footprint of sampling-dominated pipelines and enable tighter hardware-software co-design for probabilistic algorithms. The XTR-0 bridge and the simulation library accelerate experimentation and porting, lowering the barrier to test new thermodynamic algorithmic primitives and integration patterns with existing models and training loops.