top of page
Search

Inside Extropic's Thermodynamic Computing Leap: From One Chip to One Billion Pbits

Updated: Aug 12

We've written before about the Compute Economy - the idea that every economic era is defined by whatever resource is scarcest. Land constrained agriculture. Capital and machinery constrained industry. Human time constrained services. The Autonomous Economy is constrained by compute, energy, and data.


Today, one of our portfolio companies just shipped hard evidence of what solving that constraint actually looks like.


Extropic, the thermodynamic computing company we backed early, announced a full stack of updates: a new open-source programming framework, a compiler for probabilistic hardware, and its next-generation chip, Z1. Taken together, it's the clearest signal yet that the industry's brute-force approach to scaling AI - more GPUs, more power, more capital - is not the only path forward.


The problem Extropic is built to solve

Every major AI lab is currently locked in the same race: more parameters, more data centers, more electricity. It's working, but it's also running into a wall. Data centers now compete with cities for power. Hyperscalers are proposing space-based compute clusters to escape terrestrial energy limits. Robots and AR devices can't fit enough compute density into a form factor anyone would actually wear or deploy.


Extropic's founding bet is that brute-forcing today's deterministic, digital hardware to unfathomable scale can't be the endgame for algorithms that are fundamentally probabilistic. Generative AI is, at its core, sampling from distributions. So why run it on hardware built to eliminate randomness, when you could build hardware that harnesses it instead?


That's the idea behind the pbit - a probabilistic bit that uses the natural thermal noise in a transistor instead of fighting it. It's a page borrowed directly from biological intelligence, which runs on noisy, energy-efficient wetware rather than deterministic logic gates.


What Extropic Just Shipped in Thermodynamic Computing

Last fall, Extropic proved the concept with X0, its first thermodynamic chip, and put early test kits (XTR-0) into developers' hands. This week's announcement is the jump from proof-of-concept to a real stack:



  • Torx - a new open-source framework for building and training stochastic programs with gradient descent, the same way developers train neural networks today. Extropic is positioning it as the PyTorch of probabilistic computing: hardware-agnostic, and designed to run on everything from today's XTR-0 to next-generation chips. Docs and code are live at torx.ai.


  • Thermalizers - a compilation layer that takes a general Torx program and maps it onto thermodynamic silicon automatically, without a developer needing to manage low-level hardware details. For the right workloads, Extropic says this can unlock up to 10,000x better energy efficiency than GPUs.


  • Z1 - Extropic's next chip, packing over 269,000 pbits with 16-neighbor connectivity and a sampling rate above 50 MHz, all inside a die smaller than 12mm per side that draws under a single watt. Z1 will ship in two form factors: an M.2 "thermo compute stick" for edge applications like robotics and AR, and a PCIe accelerator card carrying over 4 million pbits for servers and workstations. Extropic plans to use these cards to build the world's first billion-pbit cluster in the next year.


  • An open-access API at extropic.dev that lets any developer run Torx and THRML programs on GPU simulators today, with Z1 cluster access to follow.


Extropic thermodynamic computing stack — Torx, Thermalizers, and Z1

The $75 million signal

The same week, Extropic also signed a non-binding letter of intent for up to $75 million with the U.S. Department of Commerce's CHIPS R&D Office to scale and onshore manufacturing of its thermodynamic sampling units. It's a meaningful vote of confidence from outside the venture world: the U.S. government treating energy-efficient, domestically-made AI silicon as a national competitiveness issue, not just a startup's roadmap slide.


For a company we backed at the earliest stage, having both the technical roadmap and a federal manufacturing partner line up in the same week is the kind of proof point every investor hopes for and few get this early.


Why this matters for the Autonomous Economy

We don't invest in hardware because we think chips are interesting in isolation. We invest in infrastructure that removes a binding constraint on the Autonomous Economy - the shift from AI that assists work to AI that executes it end-to-end.


That shift doesn't happen on today's compute economics. Running billions of autonomous agents and physical robots continuously, in the field, at the edge, requires an order-of-magnitude improvement in intelligence-per-watt. Extropic's approach - trading brute-force determinism for nature's own randomness - is one of the few credible paths we've seen to actually get there, rather than just scaling the current paradigm further into an energy wall.


This is also a case study in what we look for across all twelve frontiers of the Autonomous Economy: founders solving a structural constraint before the category has consensus, not building a thinner wrapper on someone else's model. Guillaume Verdon and the Extropic team weren't chasing the current compute paradigm faster - they rebuilt computation from the electrons up.


Why we backed Extropic early

We invested in Extropic before thermodynamic computing was a category anyone outside a handful of physicists and AI researchers was tracking. That's consistent with how we operate: we look for technical founders building infrastructure for a shift we believe is inevitable, well before the market prices it in.


Today's announcement - a working programming stack, a next-generation chip taped out and specced, and a nine-figure signal of confidence from the federal government - is exactly the kind of derisking milestone we underwrite for at the earliest stage. It doesn't mean the work is done. It means the thesis is playing out on schedule.


Frequently asked questions

What is thermodynamic computing?

Thermodynamic computing is an approach to hardware design that uses the natural, inherent randomness (thermal noise) present in electronic components as a computational resource, rather than suppressing it to force deterministic behavior. It's designed specifically to run probabilistic AI workloads - like sampling and generative models - more efficiently than conventional digital chips.

A pbit, or probabilistic bit, is the core building block of Extropic's chips. Unlike a conventional bit, which is forced into a fixed 0 or 1 state, a pbit fluctuates naturally and is used to represent and sample from probability distributions directly in hardware.

Extropic released Torx, an open-source framework for building and training stochastic programs; Thermalizers, a compiler that maps those programs onto thermodynamic hardware; and specs for Z1, its next-generation chip with over 269,000 pbits. It also opened public access to a simulator API and disclosed a $75 million letter of intent with the U.S. Department of Commerce to onshore manufacturing.

For the right probabilistic workloads, Extropic states its Thermalizers compilation approach can deliver up to 10,000x greater energy efficiency than GPUs. Actual gains will vary by workload and will become clearer as Z1 systems reach developers.

We invest in infrastructure that removes a binding constraint on the Autonomous Economy. Compute and energy efficiency are the current constraint on scaling AI agents and robotics, and Extropic's thermodynamic approach is one of the most credible technical paths we've seen to solving it at the hardware layer..

It's Untapped Ventures' thesis for the era where AI moves from assisting human work to executing it end-to-end - digital agents and physical robots transacting and operating with limited human involvement. Read our full breakdown in The Compute Economy and the Autonomous Economy Manifesto.


Where to go next


We're backing the infrastructure layer of the Autonomous Economy before it becomes consensus. If you're building at the compute, energy, or hardware layer of AI, pitch us.


If you want the full framework behind calls like this one, read our Autonomous Economy Manifesto.


 
 
 

Comments

bottom of page