On July 29, the U.S. Department of Commerce published a letter of intent (LOI) indicating plans for research and development incentives of up to $75 million for Extropic, a company developing semiconductors that use thermal fluctuations for computation. The target is a probabilistic computing chip called the "Thermodynamic Sampling Unit (TSU)." If the funding is finalized, development will proceed from validating prototype chips to launching a Z1 cluster and then manufacturing the successor Z1.5 at a domestic U.S. foundry.
This deal is not a purchase of a product that will immediately replace GPUs. Extropic has been validating probabilistic circuits using the small-scale X0 prototype chip and the research platform XTR-0. Z1 is the next-stage, production-scale TSU that follows this work, and the company plans to demonstrate chip performance using generative AI benchmarks. What the technical evaluation needs to confirm is not merely the power efficiency of the chip alone, but the extent to which the full computing system—including models, control, and interconnects—can be demonstrated.
Conditions Set by the LOI for Up to $75 Million
The LOI published by the U.S. Department of Commerce on the same day covers seven companies, with a combined ceiling amount of $874 million. Extropic's allocation is up to $75 million, and the department described the TSU as a technology that uses natural thermal fluctuations to probabilistically solve complex problems in simulation, optimization, and AI.
However, the LOI is not a final award. The Department of Commerce states that additional due diligence and approval will be conducted before final disbursement to each company. As a condition for receiving the final award, the government also stipulated that it would obtain a minority, non-controlling equity stake in each company. Therefore, the $75 million is not a grant already received, but rather a ceiling amount contingent on contracts and development milestones.
This point distinguishes today's announcement from ordinary fundraising. The scope of federal government backing extends to semiconductor design, prototyping of probabilistic computing as a new computational paradigm, and the manufacturing pathway through the U.S. supply chain.
Using Noise for Computation Rather Than Treating It as Noise
The starting point of the TSU is the thermal fluctuation of transistors—something that digital circuits typically try to suppress. Extropic's p-bit probabilistically toggles between two states, corresponding to output voltages of 0 and 1, with a control voltage adjusting the likelihood of each state. It is a device akin to a biased coin flip realized in circuitry.
A TSU connects a large number of p-bits with short-distance wiring to obtain samples from a probability distribution represented by an energy-based model. Each cell receives the states of neighboring cells and probabilistically updates its own next state. Extropic uses Gibbs sampling for this iterative process. Whereas GPUs mainly perform large volumes of matrix operations to support probabilistic generation and inference, the TSU is designed to make sampling itself the circuit's job.
This approach allows for increased local communication without repeatedly sending data back and forth over long distances between compute circuits and memory. The power efficiency Extropic is aiming for arises from suppressing the cost of generating randomness and data movement, and from tailoring the circuit to processes that handle probability distributions. Not all AI processing will see the same benefits; this approach presupposes models that can be centered around probabilistic sampling.
Not a GPU Replacement, But a Shift in Sampling Premises
Extropic's research team has proposed a "Denoising Thermodynamic Model (DTM)" that combines a stepwise generative process resembling diffusion models with an energy-based model. Instead of handling a complex distribution with one massive model, it breaks the process into multiple stages that go from noise back to data. The aim is to simplify the sampling at each stage and leverage the strengths of probabilistic circuits.
The roughly 10,000-fold energy efficiency figure presented in the company's paper needs to be read carefully here. It is a value from a system-level analysis that simulated a future TSU architecture on GPUs and estimated energy consumption from a physical model, projecting performance parity with GPUs on simple image benchmarks. It is not a measured value from an actual Z1 machine running large-scale generative AI, LLMs, or production clusters.
Even so, there is value in moving from GPU simulations that mimic a future TSU architecture and estimates based on physical models toward actual hardware clusters and generative AI benchmarks. What needs to be demonstrated with Z1 is not simply whether p-bits operate efficiently. It is whether the overall system loss—including image quality or generation quality, power per sample, and inter-chip communication—can be shown under equivalent comparison conditions.
Testing Scope Beyond the Z1 Chip
Under Extropic's plan, the R&D based on the LOI will cover bringing the Z1 cluster online, the infrastructure for controlling TSUs, the networking connecting chips together, and the mechanisms for rack-scale integration with existing computers. A probabilistic chip is not intended to be self-contained; it is premised on being combined with conventional computers.
On the manufacturing side, the company explains that TSUs can be built using mature CMOS processes rather than leading-edge nodes. It cites its track record of taping out and testing two variants of X0 at two separate fabs, and has laid out plans to manufacture Z1.5 at a U.S. foundry. The company has not disclosed the fab name, mass production timeline, or volume.
For this reason, the first achievement that needs to be verified is not a marketing claim of having "replaced GPUs." After the final contract is signed, the questions are: with which models will Z1 demonstrate real-machine performance, and will its power efficiency hold up even when accounting for control and communication overhead? Only once these questions are answered can probabilistic computing be evaluated as a new computing resource for data centers.