Extropic AI

Extropic AI is an American AI hardware company founded in 2022 with the stated goal of building the world's most energy-efficient computers. Rather than pursuing conventional digital accelerators, Extropic is developing thermodynamic computing hardware centered on thermodynamic sampling units (TSUs), which are inherently probabilistic devices designed to match probabilistic AI workloads. The company frames this as a fundamental shift: probabilistic software meeting probabilistic hardware. Extropic's hardware roadmap spans the XTR-0 prototype platform and the Z1 chip. According to the company, Z1 is its first production-scale chip, featuring over a quarter million pbits on a die under 12 mm a side and drawing under one watt. Extropic has said it plans to make Z1 systems available in two form factors. Alongside the silicon, the company maintains an open-source software stack, including the THRML library, Torx and Thermalizers frameworks, and an API that lets developers simulate workloads running on TSUs. Extropic was founded by a team with experience in machine learning and computer hardware from Google, IBM, Apple, and Microsoft. Its selected investors and advisors include figures such as Amjad Masad (Replit), Aidan Gomez (Cohere), Balaji Srinivasan, Naval Ravikant (AngelList), Aravind Srinivas (Perplexity), Ari Emmanuel (Endeavor), Scott Belsky (Adobe), Garry Tan (YC), and Logan Kilpatrick (Google DeepMind). The company has attracted media and podcast attention, including coverage by Wired Magazine and appearances on the Lex Fridman Podcast and Lightspeed Pod. Extropic positions itself against the data center status quo, arguing that thermodynamic computing can disrupt the economics of AI compute. Its public communications emphasize energy efficiency as the core differentiator, with the Z1 chip and TSU architecture presented as the path from prototype to production-scale systems.

Extropic was founded in 2022 with the mission of building the world's most energy-efficient computers. The founding team and early employees brought experience in machine learning and computer hardware from Google, IBM, Apple, and Microsoft. From the outset, the company pursued a non-conventional computing paradigm: thermodynamic computing, in which hardware is inherently probabilistic and therefore suited to probabilistic AI workloads. The company's early public presence included appearances on the Lex Fridman Podcast in December 2023, Garry Tan's channel in March 2024, S3 / StoryCo in April 2024, Lightspeed Pod in July 2024, SF Deep Tech Week in August 2024, and Reason Magazine's podcast in December 2024. In March 2025, Extropic appeared at TED AI San Francisco and was covered by Wired Magazine. A Core Memory Pod appearance followed in May 2025, and Decentralized AI Day 2025 took place in March 2025. On the hardware side, Extropic developed the XTR-0 prototype platform and the X0 chip, and then announced Z1 as its first production-scale chip. According to the company, Z1 packs over a quarter million pbits on a die under 12 mm a side while drawing under one watt. Extropic stated it plans to make Z1 systems available in two form factors. In October 2025, the company released a launch video titled "Thermodynamic Computing: From 0 to 1" and was the subject of coverage suggesting it aims to disrupt the data center boom. By summer 2026, Extropic published an update titled "From One to One Billion," describing progress on Torx, Thermalizers, and Z1. The company also published technical writing including "Z1T: Transformer-like models for Z1" (September 2026), "Thermalizing Stochastic Programs" (August 2026), and "A Framework for Stochastic Differentiable Programming" (August 2026). Its software stack includes the open-source THRML library, the Torx and Thermalizers frameworks, and an API for simulating workloads on TSUs. Extropic's selected investors and advisors include Amjad Masad, Aidan Gomez, Balaji Srinivasan, Naval Ravikant, Aravind Srinivas, Ari Emmanuel, Scott Belsky, Garry Tan, and Logan Kilpatrick.

• Extropic founded to build the world's most energy-efficient computers.

• Appearance on Lex Fridman Podcast.

• Appearance on Garry Tan's channel.

• Appearance at S3 / StoryCo.

• Appearance on Lightspeed Pod.

• Appearance at SF Deep Tech Week.

• Appearance on Reason Magazine Pod.

• Appearance at TED AI San Francisco.

• Coverage in Wired Magazine and Decentralized AI Day 2025.

• Appearance on Core Memory Pod.

• Coverage: Extropic Aims to Disrupt the Data Center Bonanza.

• Launch video: Thermodynamic Computing: From 0 to 1.

• Summer 2026 update: From One to One Billion — Torx, Thermalizers, and Z1; also published Thermalizing Stochastic Programs and A Framework for Stochastic Differentiable Programming.

• Published Z1T: Transformer-like models for Z1.

Developer of thermodynamic computing hardware positioned as radically more energy efficient than GPUs for probabilistic AI workloads.

• Thermodynamic sampling units (TSUs)

  Inherently probabilistic hardware devices designed for probabilistic AI workloads.

• XTR-0

  Prototype platform for thermodynamic computing.

• X0

  Early chip in the XTR-0 platform.

• Z1

  First production-scale chip: over a quarter million pbits on a die under 12 mm a side, drawing under one watt.

• THRML Library

  Open-source library for thermodynamic software development.

• Torx

  Open-source framework for programming stochastic and thermodynamic devices.

• Thermalizers

  Open-source framework for thermodynamic software.

1. 1Extropic | Homeextropic.ai
2. 2About | Extropicextropic.ai
3. 3Extropic AIwikidata.org