What Extropic AI does

Extropic AI (legal entity: Extropic Corp.) is a US-based hardware startup founded in 2022 by Guillaume Verdon (CEO) and Trevor McCourt (CTO), with a team of 11-50 physicists and engineers drawn from Google, IBM, Apple, and Microsoft. The company is developing Thermodynamic Sampling Units (TSUs), a novel class of probabilistic computer that performs sampling from programmable probability distributions using out-of-equilibrium thermodynamics and self-organization principles, rather than the deterministic matrix multiplication underpinning GPUs. The architecture is built from all-transistor probabilistic bits (pbits) fabricated in standard CMOS, enabling claimed orders-of-magnitude energy efficiency gains — including a stated potential 10,000x improvement for generative AI workloads running the company's proprietary Denoising Thermodynamic Model (DTM) algorithm. The product roadmap includes the X0 silicon prototype (proven Q1 2025), the XTR-0 development platform (released Q3 2025 for early partner beta testing), and the Z1 production chip targeted for early access in 2026, alongside the open-source THRML Python library for thermodynamic algorithm development.

The company is pre-revenue and pre-commercial, currently generating no disclosed revenue and shipping no production hardware. GTM is hybrid: enterprise field sales via an algorithmic partnership program targeting organizations with large probabilistic workloads (diffusion models, scientific simulations, world models, optimization, Bayesian filtering), combined with product-led growth through open-source software, technical content marketing, podcast appearances (Lex Fridman, Lightspeed, Reason), academic publications (arXiv), research grants to PhD students, and a newsletter funnel. The company raised a $14.1M seed round in December 2023 led by Kindred Ventures (Steve Jang) with participation from Valor Equity Partners, Buckley Ventures, HOF Capital, Julian Capital, Marque Ventures, OSS Capital, and Weekend Fund, followed by smaller follow-on rounds including approximately $920,543 from Gaingels in February 2026. Total disclosed funding is approximately $15 million. Notable advisors include Naval Ravikant, Balaji Srinivasan, Garry Tan, Aidan Gomez (Cohere), Aravind Srinivas (Perplexity), Amjad Masad (Replit), Scott Belsky (Adobe), Ari Emmanuel (Endeavor), and Logan Kilpatrick (Google DeepMind).

The core technology bet is that probabilistic AI workloads — including generative AI, scientific simulation, and Bayesian inference — are poorly served by GPU architectures originally designed for deterministic graphics rendering, and that thermodynamic computing can remove energy as the binding constraint on AI scaling. The commercial thesis depends on the Z1 chip and DTM algorithm delivering on stated efficiency claims at production scale, securing enterprise design wins before competitors in the emerging probabilistic / thermodynamic computing space.

Name

Extropic AI

Legal name

Extropic Corp.

Website

https://extropic.ai

Company type

Private

Founded year

2022

Operating status

Operating

Headcount range

11–50 employees

Short description

Extropic AI is a US hardware startup building Thermodynamic Sampling Units (TSUs), novel probabilistic chips using standard CMOS that aim to run generative AI workloads orders of magnitude more energy-efficiently than GPUs, targeting organizations with large probabilistic computing needs.

Ownership category

akta.pro rank

Product category

AI Accelerator Hardware

akta.pro primary industry

Model Compression & Optimization (Quantization, Distillation, Pruning) (HDAAACAL)

akta.pro secondary industries

AI Accelerators (GPUs/TPUs/NPUs/ASICs) (HDAAAAAA), AI Server Systems & HGX/Accelerator Platforms (HDAAAAAB)

Keywords

• Thermodynamic computing hardware
• Probabilistic AI chips
• Energy efficient AI
• Sampling accelerators
• Generative AI hardware

Headquarters

HQ city

Austin

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Personnel, Supply Chain, Marketing or Sales, Operations

Revenue model

1. Hardware Sales: Sale of thermodynamic computing hardware (TSUs/probabilistic computers) to organizations running probabilistic AI workloads. Z1 production chip targeted for early access in 2026.

Go-to-market motion2 records

Distribution channels1 record

Marketing channels6 records

Core offering

Extropic AI designs and builds thermodynamic computing hardware — Thermodynamic Sampling Units (TSUs) — that perform probabilistic sampling for generative AI and simulation workloads using orders of magnitude less energy than GPUs. The company offers a hardware roadmap (X0 silicon prototype, XTR-0 development platform, Z1 production chip), an open-source thrml Python library for algorithm development, and the proprietary Denoising Thermodynamic Model (DTM) generative AI algorithm.

Product overview

Extropic AI is building a thermodynamic computing platform centered on Thermodynamic Sampling Units (TSUs) - novel hardware that performs probabilistic sampling operations using orders of magnitude less energy than GPUs. The platform consists of hardware products (X0 prototype, XTR-0 development platform, Z1 production chip) and the thrml open-source Python library for algorithm development. The Denoising Thermodynamic Model (DTM) is the flagship generative AI algorithm designed for TSUs, capable of achieving 10,000x energy efficiency for image generation workloads compared to GPU-based diffusion models.

Differentiator

Problem solved

Functional benefit

Products and services

• Thermodynamic Sampling Unit (TSU) A new category of computing hardware that performs sampling operations from programmable probability distributions using thermodynamic principles and Gibbs sampling, instead of deterministic matrix multiplication. TSUs are composed of massive arrays of probabilistic-bit (pbit) circuits built from standard CMOS transistors and are designed for organizations running probabilistic AI workloads such as diffusion models, scientific simulations, world models, optimization, and Bayesian inference.
• Z1 Production Chip Production-scale thermodynamic computing chip representing Extropic's first commercially-targeted TSU, integrating hundreds of thousands of probabilistic-bit (pbit) circuits per chip and millions per card. Built using standard CMOS processes for mass manufacturing and targeted at enterprise customers running large probabilistic AI workloads.
• XTR-0 Development Platform Experimental testing and research platform that enables development of ultra-efficient AI algorithms by providing low-latency communication between Extropic TSU chips and traditional processors. Beta-tested by early enterprise partners as a precursor to the Z1 production chip.
• X0 Silicon Prototype First-generation silicon prototype that proved Extropic's ability to manufacture energy-efficient probabilistic circuits and generate samples from simple probability distributions. Served as the proof-of-technology milestone preceding the XTR-0 development platform and Z1 production chip.
• thrml (Open-Source Python Library) Open-source Python library enabling developers to construct thermodynamic hypergraphical models and simulate running them on Extropic's Thermodynamic Sampling Units. Built on JAX for accelerated simulations of Probabilistic Graphical Models and Energy-Based Models, and serves as the developer ecosystem anchor for Extropic's TSU hardware.
• Denoising Thermodynamic Model (DTM) Generative AI algorithm inspired by diffusion models, co-designed with Extropic's TSU hardware. DTMs generate data by gradually pulling it from noise over several steps and are reported to be up to 10,000x more energy-efficient than equivalent GPU-based diffusion models on small-scale benchmarks.

Quantifiable outcome

• 10,000x potential energy efficiency improvement for running generative AI workloads on TSUs compared to GPUs

Segments1 record

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability2 records

Feature5 records

Scale indicators2 records

Recent moves6 records

Expansion highlights6 records

Direct peers

• Lightmatter: Photonics-based AI accelerator startup pursuing radical energy efficiency for AI compute via silicon photonics. Direct peer in the novel-AI-hardware-for-energy-efficiency category.
• Graphcore: AI accelerator startup (acquired by SoftBank) building Intelligence Processing Units (IPUs) for AI workloads. Direct competitor in the custom AI chip category targeting efficiency for novel AI paradigms.
• Tenstorrent: AI accelerator startup led by Jim Keller building RISC-V based AI chips for training and inference. Competes in the custom-AI-silicon market with a focus on open architecture and efficiency.
• Cerebras Systems: AI accelerator startup building wafer-scale chips for training and inference of large models. Like Extropic, Cerebras pursues novel AI hardware architectures targeting efficiency and scale, serving enterprise AI workloads.
• Groq: AI inference accelerator company with custom LPU silicon targeting ultra-low latency and high efficiency for generative AI. Direct competitor in the AI accelerator category with a similar focus on novel hardware for AI workloads.
• SambaNova Systems: AI hardware startup building reconfigurable dataflow units (RDUs) for enterprise AI training and inference. Competes directly with Extropic in the custom-AI-chip market for large model workloads.

Broad incumbents

• IBM Research (probabilistic/quantum): Long-running research programs in probabilistic and quantum computing with large enterprise research teams. Indirect competitor for thermodynamic-adjacent computing paradigms and a potential acquirer or partner.
• NVIDIA: Dominant AI accelerator incumbent with CUDA software stack and GPU roadmap that Extropic must displace. Sets the efficiency, performance, and ecosystem benchmarks that any novel AI hardware must beat.

Emerging players

• Normal Computing: Early-stage startup building thermodynamic/probabilistic computing hardware, with very similar physics-based approach to Extropic. Closest direct competitor in thermodynamic computing paradigm.
• Rain AI: Neuromorphic AI compute startup building in-memory analog compute chips. Adjacent novel-AI-hardware competitor pursuing energy-efficient AI inference.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat4 records

Key risks6 records

Key highlights7 records

Customer concentration

• GitHubgithub.com/extropic-ai/thrml
• Xx.com/Extropic_AI

Compliance2 records

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds6 records

Investors13 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Extropic AI