Extropic chips use transistor noise

This points to Extropic trying to turn noise from a hardware flaw into the core computation. In a normal chip, transistor randomness is something engineers suppress so every bit flips the same way each time. Extropic is building chips where that randomness does useful work for probabilistic AI tasks, especially sampling and generation, which could cut the energy spent forcing uncertain math onto deterministic GPUs and quantum hardware.1 2 3 4

• The practical distinction from quantum computing is that Extropic is not chasing fragile qubits that need exotic control. It says its Thermodynamic Sampling Units are built with standard semiconductor processes and use ordinary transistor level thermal noise, which makes the path to manufacturing look more like a chip startup than a physics lab.2 3 4
• The workload fit matters. Extropic describes its hardware as purpose built for probabilistic AI, where models repeatedly draw samples from noise and refine them into outputs. That is a much narrower and more concrete target than general purpose computing, and it mirrors how generative models already work under the hood.1 2 5
• This sits in a broader wave of non GPU AI hardware, but with a different bet. Lightmatter uses light to move or compute data, Cerebras keeps huge models on one wafer scale chip, and Groq uses deterministic dataflow for low latency inference. Extropic is distinct in making randomness itself the primitive, not something the system tries to eliminate.5 6 7 1

If Extropic can prove these noisy circuits deliver cheap, repeatable sampling at scale, it could open a new lane of AI hardware optimized for generation and uncertainty, not just matrix math. The next step is turning that physics insight into developer usable systems and production chips that make probabilistic AI meaningfully cheaper than running the same work on GPUs.1 2 3 4