Thermodynamic Computing's First Silicon Is Back from the Fab. The Power Math Comes Next.

When OpenAI retired Sora 1 in the U.S. on 13 March 2026 and discontinued the Sora web and app experience on 26 April 2026, Faris Sbahi, CEO of Normal Computing, started using the shutdown as his pitch slide. On a 27 March 2026 theCUBE segment recorded at NYSE Wired, Sbahi cited outside reporting that Sora was running at roughly $15 million a day in operating cost against $2.1 million in cumulative revenue, and framed the gap as the boundary condition for generative AI at scale. OpenAI has not publicly stated its rationale for the Sora pullback, so the cost/revenue framing is Sbahi's, not OpenAI's. The underlying claim Sbahi is making is harder to dismiss: even the operators with the most capital, the most efficient inference stacks, and the most appetite for subsidizing a market position cannot run high-end generative video on the energy budget that current supercomputing hardware allows.

That is the wedge for thermodynamic computing, a field that until recently was easier to mock than to evaluate. As of mid-2026 it has produced the first artifacts that supercomputing readers can actually evaluate. Normal Computing's CN101, billed as the world's first thermodynamic computing ASIC, was announced as taped out on 12 August 2025, with Normal stating the chip was moving into characterization and benchmarking. Extropic, the more publicly visible of the two startups in the category, has described its XTR-0 prototype platform as beta-tested by early research partners and published an arXiv preprint co-authored with MIT's Isaac Chuang on the diffusion-model architecture that anchors its claims. ETH Zurich's Future Computing Laboratory has a probabilistic computing hackathon with Extropic scheduled for June 2026.

The pitch, set out in Normal's tape-out release, is that a class of probabilistic and generative workloads can run on hardware whose physical dynamics natively implement the sampling step, with up to 1,000 times less energy than the same workload on a GPU. The 1,000x figure is Normal's own projection on targeted workloads, not a peer-reviewed benchmark, and it is the number that should be scoped most carefully. But even a 10x improvement on the workloads thermodynamic chips target (diffusion models, Bayesian inference, Monte Carlo sampling) would change the power math for a meaningful slice of the AI factory buildout currently being designed around gigawatt grid interconnects and dedicated nuclear procurement.

What is actually built

The difference between Normal Computing and Extropic matters because they are at materially different points on the silicon curve.

Normal's CN101 implements two primitives in hardware. The first is large-scale linear-system solves carried out as the physical relaxation of the chip's state, a substrate-level approach to a workload that sits underneath much of scientific supercomputing and large-scale optimization. The second is stochastic sampling via Normal's proprietary Lattice Random Walk primitive, aimed at Bayesian inference, diffusion-model sampling, and Monte Carlo workloads. Patrick Coles, the company's Chief Scientist, framed the roadmap in the tape-out announcement:

"Our vision to scale diffusion models with our stochastic hardware starts with demonstrating key applications on CN101 this year, then achieving state-of-the-art performance on medium-scale GenAI tasks next year with CN201, and finally achieving multiple orders-of-magnitude performance improvements for large-scale GenAI with CN301 two years from now."

That is a three-chip roadmap with a customer-relevance threshold near the end of it, not a shipping product. CN101 is in the lab. It is not in a data center.

Extropic's XTR-0 is one step earlier on the curve: a prototype platform with low-latency CPU-to-chip communication designed to let early research partners develop algorithms against the new substrate. Extropic has stated the platform has been distributed to research partners, but no named third-party customer has publicly disclosed a deployed shipment. The next chip, Z1, has been referenced but not publicly shipped.

The honest summary: CN101 is back from the fab and being characterized. XTR-0 is a development platform. Both matter, but they are not equivalents.

The 1,000x claim, properly scoped

Normal Computing's tape-out release puts the projected efficiency advantage at "up to 1,000x" on targeted AI and scientific workloads versus GPUs. That number has been repeated, sometimes without qualification, across the broader trade press. The provenance worth keeping straight: this is a projection tied to Normal's own simulation and architectural modeling for a narrow class of sampling and linear-algebra workloads, not a peer-reviewed comparison against a published baseline.

The 1,000x figure is plausible for the specific class of computation thermodynamic chips target. It is meaningless as a general statement about AI hardware. Diffusion-model sampling, MCMC, and Energy-Based Model inference are the workloads where the physics maps cleanly to the chip. LLM training is not on the target list. Weather and climate modeling, nuclear stockpile simulation, and the codes that anchor classical supercomputing also sit outside scope. What thermodynamic computing is competing for, in the language of the AI factory buildout, is a workload-specific carve-out, not a replacement of the GPU-centric stack that Nvidia, AMD, and the hyperscaler internal-silicon programs have spent the last decade building.

Even with that scoping, the consequence is non-trivial. Generative video and image workloads are precisely the slice of the AI factory load where the unit economics are most exposed, as the Sora shutdown demonstrated. If a thermodynamic chip recaptures one order of magnitude of efficiency on that slice at production scale, the supercomputing power buildout designed around it is sized differently.

The science behind the slides

The cleanest piece of evidence that the Extropic line of work has a serious technical core is arXiv 2510.23972, "An efficient probabilistic hardware architecture for diffusion-like models," submitted in October 2025 and updated in December. The author list includes Isaac L. Chuang of MIT EECS, one of the most credentialed figures in quantum and probabilistic computing, alongside Guillaume Verdon and the Extropic team. That co-authorship is the credibility anchor for Extropic's technical claims, and it sits separately from Verdon's public profile as the effective-accelerationist figure who posts as "Beff Jezos." The persona is a fact about the company's public surface. The Chuang co-authorship is a fact about the science. They should not be conflated in either direction.

The broader research base under thermodynamic computing is wider than the two startups suggest. Lawrence Berkeley National Laboratory's Stephen Whitelam group published in Nature Communications this year on non-equilibrium thermodynamic neural networks. Supriyo Datta's group at Purdue has been the originating source of modern p-bit theory for more than a decade. The science is real. The commercial economics are unproven. Both can be true at once.

The sovereignty question is not the one it looks like

The brand-level story is US/Swiss: Normal Computing is New York–headquartered, Extropic is U.S.-based, and ETH Zurich's Future Computing Laboratory has a hackathon with Extropic on the calendar for June. That framing obscures more than it reveals. The underlying physics (probabilistic spin logic, stochastic magnetic tunnel junctions, p-bit hardware, Ising-machine architectures) has been a serious research line at Purdue under Supriyo Datta and Kaushik Roy, at Tohoku University under Hideo Ohno and Shinji Miwa, and at Tsinghua, Peking University, and the Chinese Academy of Sciences' Institute of Physics for years. The commercialization wrapper is US-led. The fundamental research base is global, and the Asian academic depth has not yet produced a brand-name commercial entrant.

The most underweighted piece of the picture is Korean. Samsung Catalyst led the $50 million strategic round in Normal Computing announced on 25 March 2026, the first major Asian-strategic capital movement into a US thermodynamic computing company. It pairs naturally with Samsung's parallel bet on vertical integration in silicon photonics, another emerging energy-efficiency substrate competing for the same AI factory budget. The pattern is consistent: Samsung is taking position in the substrate-layer technologies that will determine how the next generation of supercomputing is built, before any of them have settled into a dominant architecture.

That positioning matters in the broader context of the emerging-substrate field. Chiplets moved from architecture slides to production silicon over the last 18 months. Silicon photonics was the subject of an aggressive late-2025 consolidation wave as the copper-side suppliers locked down the optical layer. Thermodynamic chips are now the next item on the substrate stack to evaluate.

The Software Ceiling sits where it always sits

The harder question for practitioners is whether anyone can actually use these chips. The programming model is not CUDA. The relevant interfaces include Energy-Based Models, which define a probability distribution by specifying its energy function and sampling from it; Markov Chain Monte Carlo primitives the hardware runs natively rather than emulating in software; and hardware-native diffusion sampling, where the chip's relaxation dynamics implement the denoising step that conventional diffusion models simulate across many GPU passes.

Energy-Based Models have been theoretically active and practically marginal for most of the last decade. The practitioner population fluent in them is small. Extropic has published thrml, an open-source Python library that lets developers simulate TSU hardware before any chip ships. Normal Computing has not publicly released a developer SDK for CN101. The simulators are arriving ahead of the chips, which is the right order. But the population of supercomputing practitioners who can target an EBM-native architecture is the binding constraint on adoption, not the silicon timeline.

This is the Software Ceiling in its purest form. The hardware is showing up before the programming model has a practitioner base broad enough to use it. The chips will work or they will not on their own physics. Whether anyone in the supercomputing community can program them is a separate question, and it is the one that determines whether thermodynamic computing earns a workload share or stays a research curiosity.

What to watch

The next twelve months should resolve more than the previous two did. CN101 characterization results are expected to produce a first measured energy comparison against GPU baselines on Normal's target workloads, though Normal has not publicly committed a release date for the data. CN201's "state-of-the-art performance on medium-scale GenAI tasks" is Normal's stated 2026 roadmap commitment, not a verified outcome. ETH Zurich's June hackathon could produce a first public corpus of developer experience with the programming model if the event runs as scheduled. Extropic's Z1 will either ship to named third parties or it will not. Samsung's Catalyst position in Normal will either lead to a fab-side strategic relationship or remain a financial bet.

The Power Question that anchors this story is not whether thermodynamic computing replaces GPUs for the supercomputing buildout. It does not. The open question is whether a workload-specific carve-out is real enough to take a slice of the energy load off the table, and whether the practitioners who would use it can find a path through the programming model in time for it to matter.