Extropic

Overview

Extropic designs thermodynamic sampling units, chips built from ordinary transistors that use the transistors’ own thermal noise to draw samples from programmable probability distributions. Generative AI runs on sampling, so the company’s pitch is that these chips can do that work on a fraction of the energy a GPU needs. Alongside the hardware it publishes an open-source stack for writing, training and compiling probabilistic programs, and it is aiming the whole package at generative AI, scientific simulation and low-power reasoning under uncertainty.

Thesis

In February the five largest cloud companies guided to between $660 billion and $690 billion of capital spending for 2026. Four of them have since raised their guidance to a combined $720 billion to $745 billion for the calendar year, up from about $448 billion spent in 2025, and new data centers increasingly go wherever power can be found. A single modern GPU draws on the order of a kilowatt. So far the industry has answered by building more generation and more data centers.

Underneath that spending sits a physical inefficiency. A digital chip holds each bit at a voltage far enough from the noise floor that thermal fluctuations almost never flip it, and it pays for that margin in energy on every operation. Generative AI needs randomness anyway. To produce an image or a sentence, a model computes a probability distribution and then draws from it, so a GPU spends energy suppressing physical noise, performs a large amount of exact arithmetic, and then runs an algorithm to manufacture artificial randomness. The brain, noisy and imprecise, gets by on roughly 20 watts.

Demand for that compute is still rising. Inference is projected to account for 75% of all AI compute by 2030, and an agentic system uses five to thirty times more compute per task than a chatbot. Verdon’s shorthand for the company’s goal is more intelligence per watt.

Hardware that samples natively could avoid much of this cost if three things hold. The devices have to be manufacturable at scale, models have to fit the hardware without losing quality, and the energy advantage has to survive once the whole system is measured. Earlier attempts to change the basis of computing, from analog to optical to quantum annealing, often cleared the first bar and stalled on the other two.

Founding Story

Guillaume Verdon, Trevor McCourt and Christopher Chamberland founded Extropic in 2022, and all three came out of quantum computing.

Verdon is a mathematical physicist who worked as a research scientist at Alphabet, where he led quantum machine learning efforts at X and at Google and helped create TensorFlow Quantum. McCourt also worked on Google’s quantum programs, and Chamberland researched quantum error correction. That background taught the founders to treat computation as a physical process to be engineered, and it gave them a realistic sense of how long exotic hardware takes to become commercially useful. The chips themselves are built from standard transistors and run at room temperature, with no quantum components.

Verdon is also known by the pseudonym Beff Jezos, under which he co-founded effective accelerationism, an online movement advocating faster technological progress. His identity was made public in late 2023, shortly before the company announced a $14.1 million seed round led by Kindred Ventures, and Extropic emerged from stealth in March 2024 with a short technical paper describing its approach.

Product

The product is a family of probabilistic chips plus an open-source stack for programming them.

Thermodynamic sampling units

The basic component is the probabilistic bit, or pbit, a small circuit of ordinary transistors run close to the noise floor so that its output flickers between zero and one while a control voltage sets the odds of each state. Extropic describes a family of such primitives and says they generate a random sample using orders of magnitude less energy than conventional methods.

Wire pbits together so that each one’s probability depends on its neighbors, and the grid behaves like a physical system settling into low-energy configurations, a structure physicists call an Ising model. Encode a probability distribution in the connection strengths and the grid’s natural fluctuations produce samples from it. Nothing is calculated step by step. A generative task is written as an energy-based model, the Thermalizers compiler maps it onto the grid, the grid settles, a sample is read out, and a conventional processor handles everything around it.

X0, XTR-0, and Z1

X0, announced in October 2025, was the first silicon and served to demonstrate the circuits. XTR-0 is a desktop development platform built around it, and the company says it manufactured dozens of these systems and shipped them to early adopters. Z1, announced in August 2026, is the first production-scale chip.

Two traits shape what Z1 can run. Each pbit connects to sixteen neighbors, while a transformer layer connects every element to every other, so models have to be redesigned for sparse connectivity. The chip is also built on mature process nodes, which keeps costs down and lets it be made in US fabs.

The software stack

Extropic shipped its software before most people could get the hardware, so researchers could start developing algorithms in simulation. THRML, open source since October 2025, describes and simulates thermodynamic models on ordinary GPUs. Torx, open-sourced in August 2026, is where developers write probabilistic programs and train them with gradient descent, the same way neural networks are trained. Thermalizers then compiles those programs down to the hardware; a paper is out and an open-source release has been announced. Until chips arrive, early users run workloads through a simulator interface that behaves like a Z1 but runs on GPUs. The newest piece, Z1T, released in September 2026, is a family of language models redesigned for Z1’s sparse connections, with open weights and a published training recipe, though none of them can yet generate text on physical hardware.

The Z1T work also tests splitting one model across GPUs and TSUs, with each handling the operations it suits. A co-processor that sits alongside installed hardware is much easier to adopt than a replacement.

Customer

Extropic has announced no commercial customers. Descriptions of its early access program mention frontier AI research labs, a weather modeling company and government-linked research institutions. As target applications the company lists generative AI, simulation of biological systems and financial markets, and probabilistic inference for robots, autonomous systems and defense sensor platforms, and it has published hardware-grounded energy estimates for a pharmaceutical workload.

The US government is lined up as both an early customer and a funder. In July 2026 Extropic signed a letter of intent with the Department of Commerce for up to $75 million through the CHIPS Research and Development Office. The money would bring the first Z1 clusters online, demonstrate performance on generative AI benchmarks, develop rack-scale systems and qualify a domestic manufacturing path for a successor chip, the Z1.5.

Market Size

If thermodynamic computing proves competitive for mainstream generative AI, its addressable market is AI compute itself. Deloitte put the market for inference-optimized chips above $50 billion in 2026, and Gartner projected $66.1 billion of spending on AI-optimized infrastructure as a service in 2027. Both figures sit beside hyperscaler capital spending guided at $660 billion to $690 billion for 2026.

The markets that suit the hardware best are much smaller. Scientific and financial workloads already written as sampling problems form a specialized segment, roughly the one quantum annealing serves, where D-Wave has operated for more than two decades on modest revenue. How big Extropic’s opportunity gets depends on how far it can move from those sampling-native niches toward general generative AI.

Competition

Extropic’s hardest competitor is the rate at which conventional computing improves. Customers will compare its chips with whatever GPUs, model compression techniques and serving software exist when the hardware ships, and those will be better than today’s.

Conventional accelerators

NVIDIA and AMD release new data center GPUs roughly once a year, each generation improving performance per watt, and between them they own the existing software ecosystem. NVIDIA reported record quarterly revenue of $68.1 billion in its fiscal fourth quarter of 2026. Behind them, a second tier of startups, tracked across the category, builds digital chips specialized for transformer inference. In 2026 Etched raised $700 million at a $21 billion valuation, SambaNova closed the first $1 billion of a round at $11 billion, and MatX raised a $500 million Series B. All of them compete with Extropic for customers, engineers and foundry capacity, and none of them asks anyone to change how models are built.

Efficiency through software

Low-precision methods store model weights in one or two bits instead of sixteen, which cuts memory and energy by large factors on hardware people already own. PrismML reported that its compressed 27-billion-parameter model keeps 98% of the original’s benchmark performance in 5.9 GB. Inference engines such as SGLang and vLLM get more useful output from installed GPUs. Each of these attacks the same energy problem with no new chips at all.

Unconventional computing

Normal Computing is pursuing thermodynamic computing by a different technical route and has announced its own silicon, which makes it the closest direct competitor. D-Wave, founded in 1999, builds quantum annealers that sample from Ising models in hardware. It is publicly traded and still has modest revenue after more than two decades, a sign of how hard it is to find paying workloads for sampling hardware. University labs have studied probabilistic bits for years, so the core idea is public, and Extropic’s edge rests on its circuit designs, its compiler and its practical experience.

Earlier hardware bets show the range of outcomes. Cerebras built a processor the size of a wafer and took roughly a decade to reach substantial revenue, even though it kept the same mathematics as GPUs and changed only the packaging. Graphcore raised about $700 million for a new AI architecture and sold to SoftBank for less than that, because a better chip without a software ecosystem loses. Mythic’s analog in-memory computing nearly shut down in 2022 after its device-level advantages shrank at the system level. Lightmatter pivoted from computing with light to connecting chips with light and prospered, since the physics worked and the market wanted a different product. Groq struggled for years and then found its market when inference took off, a case of timing rescuing a chip company.

Business Model

Extropic has disclosed neither a commercial model nor pricing. Its announcements point to sales of chips and systems, with two Z1 form factors planned, along with cloud access to hardware and government research funding. For now it offers early access to a simulator interface.

Hardware companies with differentiated products typically earn gross margins above 60%, and building on mature nodes should keep Extropic’s production costs below those of leading-edge chips. Nearer term, the planned Commerce Department funding of up to $75 million would be non-dilutive and roughly five times the company’s total disclosed equity financing. A letter of intent is not a finalized award, though, and its terms have not been published.

Traction

In October 2025 Extropic announced X0 and the XTR-0 development platform built around it, and it says it manufactured dozens of XTR-0 systems and shipped them to early adopters. Z1 followed in August 2026, with 269,568 pbits on a die under 12 mm on a side drawing less than one watt, and early access is planned for 2027.

On the software side the company has released three open-source projects, and it published a hardware architecture paper in a peer-reviewed Nature-family journal whose authors include MIT’s Isaac Chuang. Its efficiency claims are large, and none has yet been measured on shipping hardware.

Three qualifications apply to those claims. Extropic’s own benchmark shows its sparse models need about ten times more computation than a standard transformer to reach GPT-2 level quality, so the whole advantage has to come from lower energy per operation. In the current system a conventional helper chip does most of the processing and accounts for more than 95% of system energy. And because outside access to Z1 is not expected until 2027, every efficiency figure is still a projection from simulation.

The July 2026 Commerce letter of intent is the other large piece of traction, and it carries both upside and exposure. The money would not dilute shareholders and is roughly five times the company’s entire disclosed equity funding. It is also only a letter of intent, with conditions, timing and terms unpublished, and programs change. The story suits industrial policy well, since this is a new class of chip that American factories can make today without relying on the most advanced foreign fabs. That fit brings political exposure too. The company’s statement thanked the administration and the commerce secretary by name, and support identified with one administration can be withdrawn by the next.

The company had announced no commercial customers and no revenue as of September 2026.

Valuation

Extropic’s only disclosed equity financing is a $14.1 million seed round announced in December 2023 and led by Steve Jang of Kindred Ventures. Other investors included Buckley Ventures, HOF Capital, Julian Capital, Marque VC, OSS Capital, Valor Equity Partners, and Weekend Fund, along with angels including Aidan Gomez of Cohere, Amjad Masad of Replit, Arash Ferdowsi of Dropbox, Aravind Srinivas of Perplexity, Garry Tan of Y Combinator, Naval Ravikant, Balaji Srinivasan, Scott Belsky, and Tobias Lütke of Shopify. The valuation was not disclosed, and no priced round has been announced since.

Measured against that funding, the output is unusual. In 2026 conventional AI chip startups raised rounds of $230 million to $1 billion each, including SambaNova’s first close at $1 billion, Etched at $700 million, MatX at $500 million and Euclyd at over €200 million. On $14.1 million, Extropic has produced at least two chips, shipped systems and released a software stack. That points either to financing it has not disclosed or to exceptional capital efficiency, helped by the low cost of mature manufacturing nodes.

Key Opportunities

Energy as the binding constraint

If power keeps limiting AI deployment, hardware that delivers more computation per watt becomes worth buying even when it is slower. Earlier alternative architectures failed while transistor scaling was the scarce resource, and energy has since taken that place. Even a small slice of annual hyperscaler spending redirected toward more efficient computation would be a large market for a company of this size.

Government as first customer

New semiconductor categories have historically leaned on public funding and procurement in their early years, the first integrated circuits among them. Extropic’s chips can be made domestically on mature nodes, which fits US industrial policy, and the planned CHIPS funding would pay for the move from single chips to rack-scale systems without dilution, at a stage when commercial evidence is thin. Defense applications that need low-power reasoning under uncertainty could supply early revenue.

Hybrid systems alongside GPUs

If the sampling-heavy parts of generative models can move to a low-power co-processor while dense matrix operations stay on GPUs, customers could adopt the chips gradually and keep their existing infrastructure and software. Extropic has published open weights and a training recipe so outside researchers can test the approach before hardware ships, and uptake by even a few research groups would show whether hybrid architectures are practical.

Sampling-native markets

Molecular simulation, financial risk modeling and probabilistic inference for autonomous systems are already sampling problems, so they need no redesigned transformers. These markets are smaller than generative AI, and D-Wave’s history shows sampling hardware can stay confined to them for many years. They are still the likeliest source of early customers, reference deployments and reliability data.

Key Risks

The algorithm and the hardware must both succeed

Today’s AI runs on dense matrix multiplication, which Extropic’s chips do not perform. Besides building the hardware, the company has to develop model architectures that fit sparse probabilistic chips and match transformer quality. Its own results show a roughly tenfold gap in computational efficiency at equal quality, and customers are unlikely to trade output quality for energy savings. Extropic has responded by building the models and compiler itself and releasing them openly. That is a sound strategy, and it also means a team of about 25 people is doing hardware development, systems software and machine learning research at the same time.

System-level energy accounting

How efficient the pbits are says little about how efficient the whole system is. Control electronics, memory, data movement and a conventional helper processor all draw power, and in the current design the FPGA helper chip accounts for more than 95% of it. Several earlier analog and in-memory efforts showed large device-level advantages that shrank substantially once the full system was measured. Extropic has acknowledged that much of the remaining work lies beyond the chip, and the planned government funding is aimed squarely at it.

A moving conventional baseline

GPU generations, low-precision models and better serving software each improve efficiency on their own, and their effects multiply. Take three times from GPU generations, seven and a half times from low precision and two times from serving software, and the conventional stack gets roughly 45 times better with no change to how models are built. Those factors are illustrative rather than sourced, and they overlap, because a new GPU generation’s per-watt gain already includes its low-precision hardware. Even read as a range of about 20 to 45 times, the improvement shrinks a simulated 140x advantage considerably by the time the hardware meets the alternatives of its day. PrismML’s compressed model is this kind of gain arriving on existing hardware, with no new chips required.

Commercialization, funding, and founder profile

Extropic has no announced customers, a team made up mostly of scientists, and no publicly disclosed experience in high-volume chip manufacturing or enterprise hardware sales. Its largest prospective funding source is a government letter of intent that is not yet a finalized award. Verdon’s public persona has brought the company attention and prominent early backers, and it is also polarizing, which may affect hiring, customer relationships and future financing. Political alignment that helps win government support under one administration could become a liability under the next.

Summary

Extropic is betting that the energy cost of generative AI can be cut at the level of the physics. Conventional chips spend energy suppressing thermal noise and then spend more manufacturing artificial randomness, while Extropic’s thermodynamic sampling units let the noise do the work. On $14.1 million of disclosed equity, the company has produced prototype silicon, shipped development systems, open-sourced a full software stack and announced Z1, with 269,568 pbits drawing under a watt.

The central claims are still unproven on shipping hardware. The efficiency figures of up to 10,000 times on suitable workloads and 140 times on language models come from simulation. The models need about ten times more computation for equivalent quality, a conventional helper chip consumes most of the energy in the current system, physical hardware reaches outside users in 2027 at the earliest, and no customer has been announced. The nearest analogy is D-Wave, which built elegant sampling hardware and then spent two decades looking for workloads that pay.

The physics appears to work. What Extropic still has to show is a system-level energy advantage on commercially relevant tasks that is large enough, and arrives early enough, to outpace a conventional stack improving roughly 45 times on its own while Extropic builds.