Z1 by Extropic: status and timeline

Production-scale thermodynamic unit with 269k pbits for sampling workloads. Early access starts with M.2 (edge), scales to PCIe cards (servers), then billion-pbit clusters.

ProtoStage 2 of 5

Taped out in 2026. M.2 sticks, PCIe accelerators and clusters planned for early access in 2027.

Updated Jun 2026Checked 10 Oct0 updates this week

Milestones

Next · Early access M.2 Thermo Compute Sticks and PCIe cards

Z1 design finalized and sent to fab (taped out)2026Complete.
Early access M.2 Thermo Compute Sticks and PCIe cardsTarget 2027Not yet reached.
Billion-pbit cluster availableTarget Q2 2027Not yet reached.

Most important updates

Extropic announces Z1 thermodynamic sampling unit with 269k pbits; early access planned for 2027Jun 2026

Upcoming

Q4 2026CHIPS R&D funding from U.S. Department of Commerce awarded (next)
2027Early access M.2 Thermo Compute Sticks and PCIe cards
Q2 2027Billion-pbit cluster available

Current obstacles

Competing with GPU incumbencyPorting from GPUs requires new libraries, tools and developer training.
Cost-per-sample at scale10,000x energy claims need production units showing lower TCO than GPU clusters.

Physics limits

Suited for sampling, not all inferenceDiffusion models, large-scale sampling for video synthesis, and probabilistic robotics control benefit most. Transformer-based inference and deterministic tasks are not a natural fit.
Scaling requires either fabs or massive capexTo reach thousands of units per year, Extropic must either partner with fabs (competing for wafer allocation with NVIDIA, TSMC) or build its own manufacturing, which the DOC funding may support.

How it works

3 parts

Graph structure

Sparse 16-neighbor pbits

269k pbits arranged in sparse graph, each with up to 16 neighbors. Topology matches probabilistic graphical models in generative AI.

Gibbs sampling

Parallel probabilistic inference

Pbit groups execute Gibbs sampling in hardware, drawing samples without CPU intervention or memory fetches.

Scaling

Form factors from edge to servers

Z1 dies in M.2 (edge/robotics), PCIe (servers), multi-card (image/video).

Spec sheet

Spec	Z1
Probabilistic bits per Z1 die	269,568^R (reported)
Sampling rate	>50 MHz^R (reported)
Power consumption	<1 W^R (reported)
Die size	<12 mm^R (reported)
Pbits per PCIe accelerator card	>4 million^R (reported)
Pbits in full cluster	1 billion+^R (reported)

^R reported by the company

Update log

1 update

Jun 2026

Major: Extropic announces Z1 thermodynamic sampling unit with 269k pbits; early access planned for 2027PressHardware

About Extropic

The team behind Z1

Extropic

Thermodynamic computing, probabilistic bits, Z1 chip

Extropic builds thermodynamic sampling units—specialized processors that use the thermal noise of ordinary transistors to sample directly from probability distributions. The company aims to deliver orders-of-magnitude energy efficiency gains for AI workloads compared to GPUs.

Raised $14.1M seed (Dec 2023) led by Kindred Ventures, with Aidan Gomez of Cohere and others.
Signed $75M letter of intent with U.S. Department of Commerce for CHIPS R&D funding.
Thermodynamic computing targets 10,000x energy efficiency for AI sampling.

extropic.ai (opens extropic.ai)Company profile →

Founded: 20224 yrs
Headquarters: United States
Status: Private
Valuation: private
Raised: $14M1 round
Last round: Seed · $14MDec 2023
Works in: ComputingNew kinds of computers
Coverage: 2 programs · 4 updateslatest 29 Jul 2026checked 10 Oct
Lead investors: Kindred Ventures
Partners: U.S. Department of Commerce (opens commerce.gov)
Investors: Cohere (opens cohere.ai)
People: Guillaume Verdon (opens en.wikipedia.org)CEOTrevor McCourt (opens extropic.ai)CTO

Spec

Probabilistic bits per Z1 die

269,568^R (reported)

Sampling rate

>50 MHz^R (reported)

Power consumption

<1 W^R (reported)

Die size

<12 mm^R (reported)

Pbits per PCIe accelerator card

>4 million^R (reported)

Pbits in full cluster

1 billion+^R (reported)