From One to One Billion

AI is entering an era of seemingly smooth exponential acceleration. Models keep scaling and are emerging as multi-trillion-parameter behemoths with Fields Medal-level mathematical intelligence. Data centers are multiplying, stacking seemingly endless rows of GPUs while burning ever more power. AI labs are locked in a race to scale, where the only variables that seem to matter are digital compute, energy, and capital. The whole industry is past an inflection point, and the momentum feels quasi-unstoppable.

And yet, something still feels off. Our intuition tells us the story is still incomplete. Brute-force scaling today’s digital, deterministic hardware to unfathomable scale for inherently probabilistic algorithms cannot be the endgame. Compute costs have ballooned because building enormous data centers requires land and energy that are increasingly difficult to secure. The market’s insatiable appetite for intelligence keeps straining the supply chain and demanding exponential gains from the current paradigm.

Beyond data centers, physical intelligence does not have nearly enough compute density to run capable robots entirely on local compute. Space-based compute clusters are being proposed to alleviate pressure on terrestrial data centers, while relying on designs with radiators and solar panels the size of football fields. AR wearables with sufficient compute power look comical and still cannot reach a form factor dense enough to be both powerful and appealing to consumers.

In all cases, we are hitting a physical limit: the thermodynamic limits of AI scaling under the digital paradigm. If we continue with the status quo, the apparently exponential curve will bend and local progress will plateau. For artificial intelligence to fulfill its promise, we need a non-incremental gain in the density of intelligence. We need a fundamental paradigm shift in how we make silicon think.

Our very existence, that of biological general intelligence, is proof that a better way forward can exist. One that is far removed from digital computation. One that harnesses the inherent randomness in nature. And one that is far more parameter, data, and power efficient.

At Extropic, we are pioneering this paradigm. We are taking guidance from thermodynamic physics and nature to design a new form of computing, from the electrons up. By leaving conventional bits behind and embracing probabilistic primitives such as pbits, we carved a new path that promises more intelligence per watt, per square millimeter, and per second.

It has been a multi-year journey to reinvent computation for an era dominated by AI. It has been a few months since you last heard from us, and today we have updates to share across our entire thermodynamic computing stack.

Last fall, we announced X0, our first silicon thermodynamic chip, and with it the world’s first desktop probabilistic computer, XTR-0. X0 demonstrated our novel probabilistic primitives in silicon: pbits, tiny circuits that turn the thermal noise of ordinary transistors into programmable randomness, using orders of magnitude less energy to generate a sample than conventional approaches. Those primitives are the building blocks of every chip we make, current and future.

We didn’t stop at silicon. We manufactured dozens of XTR-0 systems, our desktop experimental platform built around X0, and shipped them to early adopters who are playing with our probabilistic primitives and running the first thermodynamic programs on real hardware. We also built an XTR-0 cluster for scientific and enterprise users to remotely experiment with more than a handful of primitives, forming the first thermodynamic computing proto-cloud.

XTR-0 is a reprogrammable stochastic processor that extends beyond energy-based model sampling. To let users explore this broader space, we created Torx, which we cover below.

With our fall release, we open sourced THRML, our library for thermodynamic hypergraphical models, which lets developers experiment with thermodynamic programming close to the metal, on hardware or in simulation.

We published papers applying THRML to generative modeling and drug discovery, both showing that a Z1-like chip can achieve several orders of magnitude greater energy efficiency than GPUs.

With X0, we took thermodynamic computing from zero to one. Now it’s time to turn it up from one to one billion.

Before we get to the hardware, let’s talk software.

We heard you after our last launch. “Cool alien computer, but how am I supposed to program this thing?” Fair enough. So we built new layers of our stack, inspired by the deep learning frameworks you know and love, to make thermodynamic programming feel... less alien.

Recall that deep learning conquered software because it let us write parameterized hypotheses over programs, then search that space with gradient descent. As Andrej Karpathy has argued, gradient descent can be a better programmer than you. We’re now extending that same philosophy to stochastic devices.

Today we’re releasing Torx, our new open-source framework for general stochastic differentiable programming, going beyond strictly thermodynamic programs. In classical differentiable programming, you arrange modular parameterized functions into a directed acyclic graph. Torx lets you build the same kind of graph, but the nodes are stochastic transition kernels.

Each stochastic transition kernel can have trainable parameters optimized with gradient-based methods. The kernels act as distributions over functions, which we call factors, and can be wired in series and parallel to form a directed acyclic graph. We call these Parameterized Stochastic Circuits, or Directed Factor Graphs. The result is a modular probabilistic program with depth and a natural notion of flow, trainable end-to-end with your favorite optimizer.

This framework is directly analogous to differentiable programming frameworks like PyTorch or TensorFlow, on which the foundational breakthroughs of modern deep learning were built. Torx extends that approach to deep differentiable stochastic programs, and Thermalizers will support compiling these abstract programs onto thermodynamic hardware.

To put it simply: Torx is to THRML what PyTorch is to CUDA.

Torx is hardware agnostic. It can run on stochastic processing units of all kinds, from our very own XTR-0 today, to our next chip Z1, and beyond.

The whitepaper we’re releasing today walks through a wide array of worked examples to get you started with stochastic differentiable programming. It even features experimental results from implementations of Torx programs on XTR-0.

To get started, explore the docs, code, and examples at docs.torx.ai and on GitHub.

A framework for writing stochastic programs raises the obvious next question: how do those programs actually reach the hardware? That’s the job of Thermalizers, our upcoming thermodynamic compilation library, whose whitepaper we’re also releasing today.

Thermalizers compiles arbitrary Torx programs onto thermodynamic devices like Z1, bridging Torx and THRML and completing our thermodynamic programming stack. It works by variationally compiling each stochastic transition kernel in your program into hardware-aware energy-based models over input and output variables. The result is a directed graph of THRML models that maps directly onto our hardware, executing your program via in-memory, pipeline-parallel sampling. The whitepaper also introduces training refinements that keep compilation error under control as programs get deep.

What this means in practice: anyone can take a general, high-level stochastic program and map it onto thermodynamic hardware without ever touching low-level device details. For the right workloads, this can deliver up to 10,000x greater energy efficiency than GPUs.

Read the Thermalizers paper on arXiv, and watch our GitHub for the open-source release in the coming weeks.

Scaled up probabilistic software ate the world. Now it’s time for probabilistic hardware to scale. Three years ago we set out to scale our systems 1000x year over year.

Today we can finally share more about our progress on this journey. Introducing our newest generation of systems, based on a new chip: Z1.

Z1 comprises over a quarter million pbits arranged in a sparse graph. It samples from an energy-based model known in the literature as an Ising model, and each pbit is directly connected to sixteen of its neighbors. The chip executes a chromatic Gibbs sampling algorithm in-situ to generate samples from the thermal equilibrium state of that Ising model, at a sampling rate of over 50 MHz. Z1 fits 269,000 pbits on a chip that consumes less than one watt and measures less than twelve millimeters per side. That density is possible because a TSU stores and processes information in place, greatly lowering the energetic losses that are usually due to the long-range communication needed to shuttle bits to memory and back in traditional Von Neumann architectures.

In order for Z1 to be applicable both at the edge and in the datacenter, we plan to make Z1 systems available in several form factors. First, an edge-ready Z1 thermo compute stick: two Z1 chips, over half a million pbits, fitting in an M.2 stick form factor. We aim to enable low-power applications at the edge such as next-generation augmented reality and physical intelligence of all kinds.

For the heaviest probabilistic workloads, we’re also building Z1 accelerator cards. These pack over 4 million pbits each via 16 Z1 chips, packaged in a standard PCIe form factor for workstations and servers. These cards can be slotted right alongside your other favorite accelerators and open up early explorations of AI workloads on heterogeneous thermo-classical compute.

If scaling deep learning over the past several years has taught us anything, it’s that there are many emergent properties of differentiable programs that can only be observed empirically when systems reach sufficient scale. So to enable Thermo AI algorithms research at greater scale, we are building the world’s first large-scale thermodynamic computing cluster in 2027.

This gets us from one to one billion: the world’s first billion-pbit cluster will be built with Z1 accelerator cards.

Now that Z1 is finally taped out, we plan on making Z1 cards, sticks, and clusters ready for early access in 2027.

The world urgently needs more energy-efficient AI. This requires energy-efficient hardware and a great algorithmic migration to this new substrate.

To address this, we’re building the thermodynamic intelligence stack from the ground up, with haste. Today, we’ve open sourced more of our software tools to catalyze collaboration with the broader AI community, and hasten the migration of models to the more power-efficient thermodynamic computing substrate.

Developers can start experimenting with these new forms of programming with Torx and THRML, and start designing stochastic and thermodynamic algorithms today, to be ready for the devices to be widely available tomorrow.

To further accelerate algorithmic progress, we’re opening up an early-access simulator API that runs Torx and THRML programs on high-performance GPU simulators. This will enable you and your favorite agents to iterate faster on your thermo AI research ideas.

Over time, this API will grow to include higher-performance digital hardware accelerators, more tools for agentic research, and eventually, access to our first Z1 clusters.

If you are a developer or hacker who wants to be part of the next cycle of AI research, now is the time to join our rebellion.

We are in the foothills of thermodynamic general intelligence.

We want you to come climb this next peak with us.