Extropic logo
Academic Research · Api Tools

Extropic

Extropic builds thermodynamic computing hardware: energy-efficient probabilistic chips called TSUs. Only the open source THRML and TORX frameworks and a simulation API are accessible today; the chips themselves are announced, not sold.

Active GDPR compliant Free plan Free API available Verified by Guidaio
Overview

What is Extropic?

Extropic is a United States company building thermodynamic computing hardware, presented as radically more energy efficient than GPUs. Its core device is the thermodynamic sampling unit (TSU), described as inherently probabilistic and therefore a natural fit for probabilistic AI workloads. Two pieces of hardware are described publicly. Z1 is announced as the first production-scale chip, with over a quarter of a million pbits on a die under 12 mm a side, drawing under one watt, in two planned form factors. XTR-0 is a prototype platform. Neither is on sale, and no availability date is given. What is genuinely accessible today is software: open source frameworks for programming stochastic and thermodynamic devices, THRML and TORX, with public documentation and a public GitHub repository for THRML, plus an API that simulates workloads running on TSUs. The team is described on the About page as coming from Google, IBM, Apple and Microsoft. extropic.ai itself is a corporate site: no pricing page, no user account, no online purchase.

What it does

  • Program stochastic and thermodynamic devices through open source frameworks
  • Simulate workloads intended for TSUs through the API at extropic.dev
  • Use THRML: documentation at docs.thrml.ai, source code at github.com/extropic-ai/thrml
  • Use TORX through its public documentation at docs.torx.ai
  • Read the research papers: arXiv 2608.01615 (Thermalizing Stochastic Programs) and arXiv 2608.01612 (A Framework for Stochastic Differentiable Programming)
  • Follow the technical notes published under /writing: TSU 101, Inside X0 and XTR-0, From Zero to One, From One to One Billion
  • Review the open engineering roles listed under /about#careers
Audience

When to use Extropic / When not to

A quick filter to help you decide if Extropic is the right fit.

When to use Extropic

  • Researchers in probabilistic AI and sampling-based generative models
  • Physicists and engineers interested in thermodynamic and stochastic computing
  • Circuit designers (integrated circuit, board, analogue): the profiles Extropic is explicitly recruiting
  • Machine learning teams exploring inference at very low energy consumption
  • PhD and postdoctoral students: two arXiv papers from August 2026 plus open source libraries to build on

When not to use Extropic

  • Non-technical users looking for a ready-to-use application
  • Companies wanting to buy hardware today: Z1 is announced in the future tense and XTR-0 is a prototype
  • Anyone looking for a chatbot, an assistant or a content generator
  • Buyers who need a public price: there is no pricing page and no plan
  • Teams that need contractual support or an SLA: only contact@extropic.ai exists
Get started

How to use Extropic

A typical end-to-end flow, from setup to results.

  1. Start with the /software page for an overview of the frameworks
  2. Open the TORX documentation at docs.torx.ai/en/latest/
  3. Open the THRML documentation at docs.thrml.ai/en/latest/
  4. Pull the THRML source code from github.com/extropic-ai/thrml
  5. Reach the API through extropic.dev to simulate a workload intended for a TSU
  6. Read the two arXiv papers for the theoretical framework behind the libraries
  7. Follow /writing for hardware updates and technical notes
  8. Subscribe to the newsletter through the footer form if you want announcements
  9. Write to contact@extropic.ai for any other request
  10. Note that there is no account to create on extropic.ai, nothing to buy and no trial to start
Quick read

Pros & Cons

Pros

  • Open source frameworks and public documentation, with no account and no payment required
  • A simulation API that can be used without owning the hardware
  • Published, checkable research: two arXiv papers from August 2026
  • A technical claim backed by figures: under one watt, over 250,000 pbits on a die under 12 mm a side
  • Detailed and dated technical notes published under /writing
  • Team, investors and advisors identified on the About page

Cons

  • The hardware is not commercially available: Z1 is announced in the future tense, XTR-0 is called a prototype
  • No pricing page, no published price, no plan
  • No user account, no trial, no product demo
  • Very generic terms of service and privacy policy, scoped to community, open source and communications
  • No minimum age is stated
  • No data hosting information, no published sub-processors, no retention period
  • A single contact channel, contact@extropic.ai: no contact page and no linked LinkedIn account
Pricing

Pricing & Plans

No pricing is published on the site: there is no pricing page, no plan, no amount and no currency, so no entry price point can be quoted. What is currently accessible is free of charge and open: the THRML and TORX frameworks are open source, with a public GitHub repository and open documentation, and the simulation API at extropic.dev is presented without any stated pricing condition. For the Z1 hardware, neither a price nor an availability date has been published.

Prices and plans listed above may evolve. Always check the official pricing page before subscribing.
Trust & Privacy

Data, GDPR & hosting

A consolidated view of how Extropic handles your data.

GDPR overview

Compliance is claimed but barely documented. The privacy policy states, in a single line under Additional Considerations, that Extropic complies with international data protection regulations like GDPR and CCPA. Explicit consent is claimed for collection linked to communications and community activities, and section 5 sets out access, correction and deletion rights. Section 6 mentions that data may be transferred internationally adhering to global data protection standards, without naming a single mechanism: no standard contractual clauses, no adequacy decision. Everything else a documented GDPR programme would show is absent: no data protection officer, no dedicated privacy email address, no Article 27 representative in the EU, no published list of sub-processors, no retention period, and no effective date or last-updated date on the policy itself.

Who owns the data?

Extropic keeps ownership of everything on its side. Section 3 of the terms of service states that Extropic retains all rights to its technology, including but not limited to AI models, algorithms, and software, and grants users only a non-exclusive, non-transferable right of use, with no transfer of ownership. On personal data, the privacy policy lists names, email addresses and contact details, community interaction data, and technical data such as browser and operating system. It states that data is not shared with third parties, apart from service providers bound by confidentiality clauses and legal obligations, and that access is limited to authorised personnel. Note the scope: these documents describe a community and open source relationship, not a SaaS product.

Reuse rights

The terms of service (section 2) require use that is consistent with the intended purpose, respectful and lawful, and prohibit any misuse of Extropic's technological or computing resources. Section 3 grants a non-exclusive, non-transferable right of use with no ownership of the intellectual property, so the technology itself may not be redistributed or passed on without permission. On personal data, the privacy policy (section 5) gives users the right to request access, correction and deletion through the contact channels provided. What the documents do not cover is the reuse of any outputs or results: the case is simply not addressed, so nothing on the site authorises or restricts it. On the code side, THRML is published in a public GitHub repository (github.com/extropic-ai/thrml), but its licence was not collected, and the exact reuse terms of the code should be checked in the repository itself.

Data retention & training

Retention summary
Retention is essentially undocumented. The privacy policy sets no retention period for any category of data: no duration, no purge schedule, no criteria for deciding how long information is kept. What does exist is a deletion route, since section 5 gives users the right to request access to, correction of, or deletion of their data through the contact channels provided, and a breach commitment, since section 4 announces prompt notification in the event of a data breach. There is no mention of anonymisation or pseudonymisation, and nothing describes what happens to data once a relationship ends. In practice, deletion depends on asking for it.
Subprocessors disclosed
No

Hosting summary

The site gives no hosting location. No country and no region is named for where user data is stored. The privacy policy (section 6) mentions that data may be transferred internationally while adhering to global data protection standards, but names no destination country and no transfer mechanism. Section 4 announces secure storage and regular audits, with no infrastructure detail whatsoever: no provider, no data centre, no region. No sub-processor is named anywhere on the site. A network lookup places the website itself on an anycast node, with an A record at 75.2.60.5 belonging to AS16509 Amazon.com Inc. and geolocated in Seattle, United States. That describes how the extropic.ai site is delivered, not where user data is hosted, and the two should not be confused. Anyone with a jurisdictional requirement will have to ask contact@extropic.ai directly, because the published documents answer none of it.

Watch-outs

Things to keep in mind

Risks and trade-offs to weigh before adopting Extropic.

  • Do not confuse an announcement with availability: Z1 is described in the future tense and XTR-0 is a prototype
  • The performance figures are published by the vendor and are not verified by any third party on the site
  • Generic legal documents whose scope is limited to community, open source and communications
  • No data retention period is stated anywhere in the privacy policy
  • No minimum age is set, although the newsletter form is open to anyone
  • International data transfers are mentioned without naming any mechanism
  • Building a project on a hardware stack that is not yet available remains speculative
Setup

Setup & Integrations

Technical difficulty

High, and for a technical audience only. Nothing is installed from extropic.ai itself, which is a corporate site. Real use means integrating libraries: THRML from its public GitHub repository, TORX through its documentation, and the simulation API through extropic.dev. The profiles Extropic recruits for, covering fundamental physics, integrated circuit and board design, C++ and Python, indicate the expected level. Expect a researcher or engineer skill set: there is no consumer interface and no guided onboarding.

Deployment

API
Company

Behind Extropic

Company name
Extropic Corp.
Founded
03/12/2023
Country of origin
🇺🇸 United States
UBO
INFORMATION_NOT_FOUND
UBO country
INFORMATION_NOT_FOUND
Domain registrar country
🇺🇸 United States
Support contact
Official links

Resources

All the official URLs gathered for verification and reference.

FAQ

Frequently asked questions

What is a TSU?
A thermodynamic sampling unit. It is hardware that is inherently probabilistic, which Extropic presents as the right fit for probabilistic AI workloads rather than the deterministic arithmetic GPUs are built around.
Can I buy Extropic hardware today?
No. Z1, the first production-scale chip, is announced in the future tense in two planned form factors with no availability date, and XTR-0 is described as a prototype platform.
What is actually accessible right now?
The open source frameworks THRML and TORX, their public documentation at docs.thrml.ai and docs.torx.ai, the THRML repository on GitHub, and an API at extropic.dev that simulates workloads intended for TSUs.
What does it cost, and do I need an account?
No pricing is published: no pricing page, no plan, no amount. And no account is required. extropic.ai only offers a newsletter form and the address contact@extropic.ai.
Conclusion

Should you pick Extropic?

Extropic is a semiconductor company, not a ready-to-use AI product, and that distinction should be made before spending time on it. What can be used today is the software layer: the open source THRML and TORX frameworks, their public documentation, and an API that simulates TSU workloads. The hardware cannot be used at all: Z1 is announced in the future tense and XTR-0 is described as a prototype. The realistic audience is therefore researchers, hardware engineers and machine learning teams working on probabilistic methods, not buyers. The legal documents are thin and clearly written for an open source community rather than professional use. Progress is at least checkable, through dated arXiv papers, dated technical notes and press coverage.