Guillaume Verdon's Extropic and the $75 Million CHIPS Deal for Thermodynamic Sampling Units
Guillaume Verdon’s company Extropic signed a letter of intent with the US Department of Commerce on 29 July 2026 for up to $75 million in CHIPS research and development incentives. The money is meant to bring Extropic’s thermodynamic sampling units, chips that use thermal noise to draw random samples, from prototype to production at a US foundry. It is a letter of intent and not a final award, and the government takes a minority equity stake as a condition.
In short, the deal funds the step that a substrate independence test needs most, chips that exist in quantity and can be measured. It does not prove any of Extropic’s energy claims, and nothing in it has been measured by this project.
What the letter of intent says
The award was one of seven letters of intent that the Commerce Department announced together, worth $874 million in total under the CHIPS and Science Act. The National Institute of Standards and Technology published the list in its press release of 29 July 2026. The CHIPS Research and Development Office runs the program.
| Company | Up to | What the money is for |
|---|---|---|
| GlobalFoundries | $300 million | Co-packaged optics beside AI processors |
| Kepler | $245 million | 3D and ferroelectric AI memory |
| Multibeam Corporation | $140 million | Advanced packaging for stacked chips |
| Extropic | $75 million | Thermodynamic sampling units |
| Thintronics | $50 million | Low-loss dielectrics for interconnects |
| OBSIDIA Semiconductors | $34 million | Counterfeit component identification |
| Aeluma | $30 million | Indium-phosphide-free photonic substrates |
Extropic is the only company on the list that builds a new kind of computer. The other six improve parts of the conventional stack, such as memory, packaging, optics and materials. The release describes the Extropic award as money “to develop thermodynamic sampling units (TSUs) which use natural thermal fluctuations to probabilistically solve complex problems.”
Three conditions apply to every company on the list. The letters are not yet awards, and the release says “there will be further diligence and approval by the Department before final awards are made.” The government receives “a minority, non-controlling equity stake in each company as a condition for receiving the funds.” Extropic’s own announcement adds that the funding is non-binding and depends on definitive agreements and project milestones.
What a thermodynamic sampling unit is
A thermodynamic sampling unit is a chip made of p-bits. A p-bit is a binary unit that flips at random, and its bias is set by its inputs. When many p-bits are coupled, the grid settles into states with the probabilities of an Ising model, so the chip produces the samples that a GPU has to compute step by step. Extropic builds its p-bits from standard CMOS transistors, which is why it says the chips can be made on mature process nodes.
The company names four chips in its announcement, Thermodynamic computing chips in America.
| Chip | Role, as Extropic states it | Status on 29 July 2026 |
|---|---|---|
| X0 | Prototype silicon. Two variants taped out and tested at two different fabs | Exists, shipped to early users |
| Z1 | First production-scale TSU | Taped out |
| Z1 clusters | First systems for benchmarks on generative AI workloads | To be funded by the deal |
| Z1.5 | An iteration of Z1 “fabricated at a U.S. foundry” | Planned, no date given |
The announcement names no foundry and no process node, and it gives no date for Z1.5. Z1’s own figures, about 270,000 p-bits in eight cores, are set out in this site’s post on Extropic THRML, the open source thermodynamic sampler. They remain vendor figures.
Verdon framed the deal in one sentence. “Brute-forcing the digital, deterministic paradigm to unfathomable scale cannot be the endgame.”
The energy claim and its status
Extropic says its chips reach “energy-efficiency gains of orders of magnitude over conventional GPUs” on key generative AI workloads. The announcement gives no number. The figure most often quoted, about 10,000 times less energy, comes from a simulation in a 2025 paper by Andraž Jelinčič, Trevor McCourt, Guillaume Verdon and co-authors (arXiv:2510.23972). It is an estimate for a simple image benchmark. No independent group has published a measurement on a Z1 chip.
The deal changes one thing about that status. It pays for Z1 clusters and for benchmarks on generative AI workloads, so measurements should follow. A measured result from a production chip would be the first data that an outside group could check against the simulation.
What this means for the consciousness question
This site studies consciousness as an emergent property under functionalist emergentism. That view predicts that a system’s causal organization carries its properties, whatever material runs it. A thermodynamic chip is one of the hardest tests of that prediction. It replaces deterministic arithmetic with physical noise, so any property that survives the move cannot depend on exact digital computation.
The test needs three things. It needs chips that exist in numbers, an interface that a research group can program, and measurements that someone other than the vendor can repeat. Before this deal, only the first prototypes existed. If the funded Z1 clusters come online and Extropic opens access as it has with X0, the first two conditions could be met in 2027.
The opposing view is also on record. Todd Feinberg and Jon Mallatt, whose six neurobiological features shape this project’s design, argue that the physical substrate is critical for sentience. On their view a move to p-bits could change more than speed and cost. The question stays open in both directions until a measured result exists.
Comparison to The Consciousness AI
The project runs p-bits only in software. Its emulator, models/thermodynamic/p_bit_emulator.py, implements the same chromatic block Gibbs sampling on a 16-neighbour grid that Z1 is stated to run, and an optional wrapper calls Extropic’s THRML library in software. The repository holds no driver for any Extropic chip. The project’s own test of whether its representations survive a noisy p-bit channel gave a narrow result. Known stimuli stay distinguishable, but shape decoding falls to chance under sensory noise. No result from the project shows that its architecture is substrate independent.
The Thermodynamic Computing hub shows the two systems in the browser, an Ising lattice near its critical temperature and a ring of Kuramoto oscillators. Readers who want to run p-bit sampling on physical hardware before Z1 is available can use an FPGA board or a USB noise source, and the thermodynamic computing hardware guide lists boards that are sold today. None of them is an Extropic chip.
The energy side of the question, how much power a growing intelligence needs and where it comes from, is set out in the Kardashev scale read as an energy budget. The wider effort to agree criteria for consciousness in any system is in the race to define AI consciousness.
What is known and what is not
| Known on 29 October 2026 | Not known |
|---|---|
| Extropic signed a letter of intent for up to $75 million in CHIPS R&D incentives on 29 July 2026 | Whether the final award is made, and on which terms |
| The government takes a minority, non-controlling equity stake | The size of the stake |
| X0 exists in two variants from two fabs, and Z1 is taped out | Which US foundry and process node Z1.5 uses, and when |
| Extropic states orders of magnitude energy gains on generative AI workloads | Any independent measurement of a Z1 chip |
| The project emulates the same sampling scheme in software | Whether any property of the project’s architecture survives on p-bits |
Guillaume Verdon’s Extropic now has a funded path from prototype to US production. The thermodynamic sampling unit remains a vendor claim until the Z1 clusters run and someone outside the company measures them.