Extropic is a thermodynamic computing startup building probabilistic silicon as an alternative to GPUs for AI workloads. This launch introduces Z1T, its first family of transformer-like models designed to run on Z1, a production-scale chip featuring hundreds of thousands of probabilistic circuits per die and manufacturable on standard CMOS processes . The pitch to founders and researchers watching AI infrastructure costs is direct: up to 140x energy efficiency gains over GPUs, along with a new scaling law for sparse transformers .
Rather than forcing transformers onto probabilistic hardware unchanged, Extropic draws on the Boltzmann-machine lineage to build sparse neural layers that map to Z1's graph of pbits, then pairs the chip with an on-board FPGA to handle the rest of the decode pipeline. The team is releasing this as an open study, with open weights on Hugging Face and a training recipe on GitHub , and frames Z1T as early co-design work meant to inform future chip generations. It matters now because inference energy is becoming a hard constraint on model deployment, and Z1T is one of the first concrete attempts to show a non-GPU substrate can run transformer-class workloads.
Extropic was founded by Guillaume Verdon and Trevor McCourt, former members of Alphabet's quantum computing research team , with Verdon as CEO and McCourt as CTO, joined by principal architect Christopher Chamberland . The company raised a $14.1 million seed round and has been shipping a full stack around Z1 including a compiler and open-source programming framework for probabilistic hardware.
Comments (8)
Sign in to join the discussion.
P
Priya NandakumarSep 4, 2026
140x energy efficiency gains is a wild claim to drop on a Tuesday with two images and no follow-up thread. Where is the throughput per watt breakdown, the marketing team was clearly asleep at the wheel.
S
sam k.Sep 4, 2026
Bold to launch a chip family with a tweet that has one comma and a link. The copy reads like it was drafted in a group chat at 2am and I respect that.
K
Kostas VlachosSep 4, 2026
Sparse probabilistic hardware sounds like something my thermodynamics professor mumbled after his third ouzo. Curious how Z1T handles inference latency compared to a distilled dense baseline on the same tasks.
Y
Yael Ben-AriSep 4, 2026
Is there a paper accompanying the scaling law claim or are we citing tweets in NeurIPS submissions now. Would love to see the exponent and the range it holds over.
H
hiroSep 4, 2026
Every time I see a hardware startup post a new scaling law I remember I still haven't finished shipping my todo app. How big is the team behind Z1, feels like six people and a soldering iron.
A
Amara O.Sep 4, 2026
The blog visual with the chip glow is doing heavy lifting for a launch that has basically no benchmarks in the tweet itself. Sell me the graph, not the vibe.
R
Ren MarchettiSep 4, 2026
Congrats on inventing a transformer that runs on vibes and Boltzmann distributions. Do I still need a datacenter or can I run this off a potato battery, asking for my seed round.
D
Devon PrakashSep 4, 2026
We prototyped something adjacent to probabilistic accelerators at my previous employer in 2019 and shelved it because nobody wanted to rewrite their stack. Curious what changed in the developer ergonomics story here.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Extropic
building the ultimate substrate for intelligence
building the ultimate substrate for intelligence
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Extropic is a thermodynamic computing startup building probabilistic silicon as an alternative to GPUs for AI workloads. This launch introduces Z1T, its first family of transformer-like models designed to run on Z1, a production-scale chip featuring hundreds of thousands of probabilistic circuits per die and manufacturable on standard CMOS processes . The pitch to founders and researchers watching AI infrastructure costs is direct: up to 140x energy efficiency gains over GPUs, along with a new scaling law for sparse transformers . Rather than forcing transformers onto probabilistic hardware unchanged, Extropic draws on the Boltzmann-machine lineage to build sparse neural layers that map to Z1's graph of pbits, then pairs the chip with an on-board FPGA to handle the rest of the decode pipeline. The team is releasing this as an open study, with open weights on Hugging Face and a training recipe on GitHub , and frames Z1T as early co-design work meant to inform future chip generations. It matters now because inference energy is becoming a hard constraint on model deployment, and Z1T is one of the first concrete attempts to show a non-GPU substrate can run transformer-class workloads. Extropic was founded by Guillaume Verdon and Trevor McCourt, former members of Alphabet's quantum computing research team , with Verdon as CEO and McCourt as CTO, joined by principal architect Christopher Chamberland . The company raised a $14.1 million seed round and has been shipping a full stack around Z1 including a compiler and open-source programming framework for probabilistic hardware.
Comments (8)
140x energy efficiency gains is a wild claim to drop on a Tuesday with two images and no follow-up thread. Where is the throughput per watt breakdown, the marketing team was clearly asleep at the wheel.
Bold to launch a chip family with a tweet that has one comma and a link. The copy reads like it was drafted in a group chat at 2am and I respect that.
Sparse probabilistic hardware sounds like something my thermodynamics professor mumbled after his third ouzo. Curious how Z1T handles inference latency compared to a distilled dense baseline on the same tasks.
Is there a paper accompanying the scaling law claim or are we citing tweets in NeurIPS submissions now. Would love to see the exponent and the range it holds over.
Every time I see a hardware startup post a new scaling law I remember I still haven't finished shipping my todo app. How big is the team behind Z1, feels like six people and a soldering iron.
The blog visual with the chip glow is doing heavy lifting for a launch that has basically no benchmarks in the tweet itself. Sell me the graph, not the vibe.
Congrats on inventing a transformer that runs on vibes and Boltzmann distributions. Do I still need a datacenter or can I run this off a potato battery, asking for my seed round.
We prototyped something adjacent to probabilistic accelerators at my previous employer in 2019 and shelved it because nobody wanted to rewrite their stack. Curious what changed in the developer ergonomics story here.