{"slug": "extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs", "title": "Extropic publishes its thermodynamic computing case as AI power demand climbs", "summary": "Extropic, a startup founded by physicist Guillaume Verdon, published a case for thermodynamic computing on July 29th, arguing that AI scaling will eventually be constrained by electricity and cooling demands of conventional data centers. The company is designing probabilistic circuits that use natural transistor noise for sampling, aiming to reduce energy consumption compared to GPUs, with a 2026 paper estimating a 10,000x energy efficiency gain on a simple image benchmark. Extropic has raised a $14.1 million seed round and plans early access for its production-scale Z1 chip in 2026.", "body_md": "[Guillaume Verdon (@GillVerd)](https://x.com/GillVerd), the physicist who founded Extropic after working on quantum machine learning at Google and Alphabet's X, published [a native article on X](https://x.com/extropic/status/2082459450166968328?s=46) on July 29th as Extropic presses its case for rebuilding AI computing around thermodynamics.\n\nExtropic's wager is that AI scaling will eventually collide with the electricity and cooling requirements of conventional data centers. Verdon and co-founder Trevor McCourt are trying to address that constraint below the software layer, designing probabilistic circuits that use the natural noise inside transistors to perform sampling operations.\n\nThat approach requires Extropic to build a new hardware and software stack. Extropic's chips are designed around probabilistic bits, or p-bits, which produce tunable random binary signals. Networks of those circuits can sample directly from probability distributions used by energy-based models, a class of machine-learning systems that represents relationships through an energy function.\n\nGPUs can perform the same general category of sampling through sequences of deterministic calculations. Extropic wants the physical behavior of its circuits to do much of that work directly, reducing the data movement and arithmetic that consume energy in conventional processors.\n\n### From quantum software to noisy transistors\n\nVerdon and McCourt previously worked together on TensorFlow Quantum, software that connected quantum-computing research with Google's machine-learning framework. Verdon developed early work behind the project during his graduate research at the University of Waterloo and later served as quantum technology lead on Alphabet X's Physics and AI team. McCourt went on to research quantum control and metrology at MIT.\n\nThe pair founded Extropic in 2022 after shifting their focus from quantum systems, where noise is usually an error to suppress, toward room-temperature transistor circuits that use noise as a computational resource. Extropic raised a [$14.1 million seed round](https://www.extropic.ai/accelerate) led by Kindred Ventures in 2023. Buckley Ventures, HOF Capital, Julian Capital, Marque VC, OSS Capital, Valor Equity Partners and Weekend Fund participated alongside angel investors including Aidan Gomez, Amjad Masad, Arash Ferdowsi, Garry Tan, Naval Ravikant and Tobias Lutke.\n\nExtropic has since produced [X0](https://extropic.ai/hardware), a prototype chip intended to validate its all-transistor probabilistic circuits at room temperature. XTR-0, Extropic's development platform, connects two X0 chips to a conventional processor through low-latency interfaces. That hybrid design lets researchers test thermodynamic algorithms while leaving control and other workloads on established hardware.\n\nExtropic is also developing Z1, its planned production-scale thermodynamic sampling unit. Extropic has listed early access for 2026, though Extropic has not published customer names, pricing, production volumes or commercial performance results.\n\n### The 10,000x result remains a modeled comparison\n\nThe strongest published support for Extropic's efficiency thesis comes from a [2026 paper in npj Unconventional Computing](https://www.nature.com/articles/s44335-026-00075-3.pdf), co-authored by Extropic researchers and MIT professor Isaac Chuang.\n\nThe researchers described an all-transistor probabilistic architecture for running a new class of generative system called a denoising thermodynamic model. Their analysis found that a future device based on Extropic's architecture could match GPU-based models on a binarized Fashion-MNIST image benchmark while consuming about 10,000 times less energy per generated sample.\n\nThat figure combines measurements from Extropic's probabilistic chip with circuit models and simulations. It is an estimate for a simple image-generation benchmark, rather than a measurement from a production deployment or a comparison against a frontier language model. Extropic still has to show that the advantage survives larger chips, manufacturing variation, communication between components and workloads complex enough to support a commercial market.\n\nThe algorithm is part of the hardware strategy. Extropic's denoising thermodynamic models break generation into a sequence of simpler energy-based models, allowing the probabilistic circuits to perform repeated sampling steps. Extropic is effectively co-designing algorithms that fit its hardware instead of adapting the transformer workloads that dominate current GPU clusters.\n\nDevelopers can experiment with that approach through [THRML](https://extropic.ai/software), Extropic's open-source Python and JAX library. THRML simulates thermodynamic sampling units on conventional processors, giving researchers a way to build and train probabilistic models before production-scale Extropic hardware is available.\n\n### Extropic still has to build a market around a new primitive\n\nExtropic is entering a market where Nvidia GPUs and Google TPUs already benefit from mature compilers, extensive developer tooling and a large installed base. Alternative accelerators often struggle because customers must rewrite software for uncertain gains and because chip production requires long development cycles and substantial capital.\n\nExtropic's path is harder in one respect: its architecture is designed for probabilistic and energy-based workloads, rather than as a faster drop-in processor for today's transformer models. Commercial adoption therefore depends on Extropic proving the hardware and persuading developers that a different model architecture can deliver useful results.\n\nExtropic also faces direct work from [Normal Computing](https://normalcomputing.com/blog/normal-computing-announces-tape-out-of-worlds-first-thermodynamic-computing-chip), which has developed thermodynamic hardware for probabilistic computing. Other efforts are pursuing photonic processors, compute-in-memory chips and neuromorphic systems to lower the cost of AI inference.\n\nVerdon's bet gives Extropic a distinct position in that field. Extropic is trying to replace deterministic computation with a new sampling primitive and build the models, developer tools and silicon around it. The published results establish a technical basis for that approach. Z1 and the workloads its early users choose will determine whether thermodynamic computing can move beyond research hardware.", "url": "https://wpnews.pro/news/extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs", "canonical_source": "https://runtimewire.com/article/extropic-thermodynamic-computing-ai-energy-pitch", "published_at": "2026-07-29 15:43:36+00:00", "updated_at": "2026-07-29 15:59:34.658806+00:00", "lang": "en", "topics": ["ai-chips", "ai-infrastructure", "ai-research", "ai-startups"], "entities": ["Extropic", "Guillaume Verdon", "Trevor McCourt", "Kindred Ventures", "Google", "Alphabet X", "MIT", "Isaac Chuang"], "alternates": {"html": "https://wpnews.pro/news/extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs", "markdown": "https://wpnews.pro/news/extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs.md", "text": "https://wpnews.pro/news/extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs.txt", "jsonld": "https://wpnews.pro/news/extropic-publishes-its-thermodynamic-computing-case-as-ai-power-demand-climbs.jsonld"}}