Normal Computing
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,147 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,147 words
Add missing citations, update stale details, or suggest a clearer explanation.
Normal Computing is an American deep-tech startup developing thermodynamic computing hardware and AI-driven electronic design automation (EDA) software. Founded in 2022 in New York City by veterans of Google Brain, Google X, Palantir, and Los Alamos National Laboratory, the company is one of the most prominent commercial efforts to build a new class of "physics-based" application-specific integrated circuits (ASICs) that exploit thermal noise, stochastic dynamics, and dissipation as computational resources rather than treating them as engineering nuisances.[^1][^2][^3] Normal positions itself at the intersection of two converging crises: the runaway energy footprint of frontier AI training and the rising design cost and complexity of advanced semiconductors. The firm pursues both problems with a single thesis: pair an AI-native EDA platform with a novel "Carnot architecture" of stochastic processors, and use each side of the business to fund and accelerate the other.[^4][^5]
In August 2025, Normal Computing announced the successful tape-out of CN101, which it describes as the world's first dedicated thermodynamic computing chip, targeting linear-algebra and stochastic-sampling primitives at up to a thousandfold energy advantage over comparable GPU workloads.[^6][^7][^8] In March 2026 the company closed a $50 million strategic round led by the Samsung Catalyst Fund, bringing total disclosed funding to more than $85 million.[^4][^9] Normal sits alongside Extropic, Lightmatter, Mythic, Etched, Rain AI, and other "post-Moore" hardware startups attempting to design computing substrates whose physics is matched, rather than abstracted away from, the workloads they accelerate.[^10][^11]
Normal Computing was incorporated in 2022 in New York City. Its three co-founders are CEO Faris Sbahi, Chief Technology Officer Antonio J. Martinez, and Chief Product Officer Yui-Hong "Matthias" Tan.[^2][^12] All three had worked together at Alphabet on production probabilistic-machine-learning systems before leaving to start the company. Sbahi was the valedictorian of his class at Duke University and joined Google as a software engineer working on YouTube Ads auctions, algorithmic game theory, and production ML and data-science tooling, before becoming a research scientist at X, the Moonshot Factory, where he worked on decision-making under uncertainty and Bayesian machine learning.[^12] Martinez and Tan held parallel roles at Google Brain and X, building infrastructure for uncertainty-aware AI at Alphabet scale.[^2][^4]
The founders' stated motivation was that the production AI stacks they had helped build at Google were dominated, in real-world deployments, by problems of calibration, uncertainty, and out-of-distribution generalization. They argued that those failure modes had a common root: software written for deterministic von Neumann processors had been forced to simulate the stochastic, sample-based mathematics that probabilistic AI actually requires. Their answer was to design hardware whose native operations are sampling and inference, not deterministic floating-point arithmetic.[^1][^13]
Within months of incorporation, Normal recruited Patrick J. Coles as Chief Scientist. Coles had previously been the principal investigator of Los Alamos National Laboratory's near-term quantum computing group, where he organized the LANL Quantum Computing Summer School and was a major contributor to the literature on quantum machine learning and variational quantum algorithms.[^14] His arrival anchored a research-heavy strategy that produced a steady stream of arXiv papers throughout 2023 and 2024, several of which laid the theoretical groundwork for the company's later hardware.[^13][^15][^16]
Normal closed an $8.5 million seed round in June 2023, led by Celesta Capital and First Spark Ventures with participation from Micron Ventures and other investors.[^17][^18] In October 2024 the company was selected as one of twelve teams for the United Kingdom Advanced Research and Invention Agency (ARIA) £50 million "Scaling Compute" programme, which targets a 1,000x reduction in the unit cost of AI compute.[^3] By early 2025 Normal disclosed a further $35 million round backed by Eric Schmidt's First Spark Ventures, Celesta Capital, Drive Capital, ARIA, Micron Ventures, Samsung Next, Intel Ignite, and the National Security Innovation Network, with the company expanding into San Francisco, London, and Copenhagen.[^5] The CN101 tape-out followed in August 2025, and the Samsung Catalyst-led $50 million strategic round closed in March 2026.[^4][^6][^9]
Thermodynamic computing treats a physical substrate as a sampler from a programmable probability distribution rather than as a deterministic function evaluator. The basic idea is older than Normal Computing and traces to a series of mid-twentieth-century results on the thermodynamics of information, including Maxwell's demon, Landauer's principle, and the Crooks fluctuation theorem.[^11] Modern thermodynamic computing inherits a key insight from that lineage: in a closed physical system relaxing to equilibrium, the distribution over states is uniquely determined by a controllable potential energy function. If a designer can shape that potential to match a target probability distribution, the device produces a sample from the distribution simply by waiting for it to equilibrate. Energy that a conventional digital chip spends suppressing thermal noise is, in this regime, spent by the thermodynamic device generating useful work.[^15][^16]
Normal's chief scientist Patrick Coles, together with several co-authors at Normal and academic collaborators, formalized this picture in the February 2023 arXiv paper "Thermodynamic AI and the fluctuation frontier" (arXiv:2302.06584).[^19] The paper argues that an apparently disparate set of probabilistic AI algorithms, including generative diffusion models, Bayesian neural networks, Monte Carlo sampling, and simulated annealing, can be unified under a single mathematical framework in which the algorithm is the equilibration of a stochastic differential equation. That observation immediately suggests that a single physical substrate, properly programmed, could implement any of those algorithms natively.[^19]
Normal's first concrete device, described in its December 2023 paper "Thermodynamic Computing System for AI Applications" (arXiv:2312.04836, later published as Melanson et al., Nature Communications 16:3757, April 2025), is a printed-circuit-board prototype called a stochastic processing unit (SPU).[^13][^20] The SPU consists of eight RLC unit cells with all-to-all switched-capacitor coupling. Voltages across the capacitors evolve under Langevin dynamics; at equilibrium the joint voltage distribution is a multivariate Gaussian whose covariance is the inverse of a precision matrix programmed into the coupling network. Reading out the voltages therefore produces samples from a user-specified Gaussian, and time-averaging the samples computes the inverse of the precision matrix.[^13][^20]
The published prototype operated with a 12 MHz measurement frequency and demonstrated 8x8 matrix inversion across three nominally identical SPU copies, with experimental error attributed to component tolerances. The authors projected that an integrated-circuit version of the architecture, with the inductors replaced by purely resistive coupling, would cross GPU performance at roughly d ~ 3,000 dimensions and reach an order-of-magnitude speedup at d = 10,000, with even larger savings in energy. Crucially, the asymptotic scaling for sampling is O(d²) versus O(d³) for digital Cholesky decomposition, so the advantage grows with problem size.[^13][^20]
CN101, Normal's first silicon implementation, replaces the LC oscillators of the prototype with on-chip RC networks and integrates the random-walk generator that produces stochastic transitions on the chip itself. Normal calls the resulting design its Carnot architecture, a reference to the Carnot cycle as an ideal limit of thermodynamic efficiency.[^6][^7] The Carnot architecture is engineered around two computational primitives: dense linear-algebra and matrix operations, and a proprietary Lattice Random Walk (LRW) sampler that produces samples from a programmable stochastic process. Both primitives are realized as the natural equilibrium behavior of a configurable analog network, rather than as software loops running on top of a deterministic digital substrate.[^6][^7][^8]
Thermodynamic computing is closest in spirit to neuromorphic computing in that both reject the strict separation between memory and compute that defines the von Neumann architecture. The key contrast is that classical neuromorphic chips, including Intel's Loihi and IBM's TrueNorth, still rely on deterministic spike events whose stochasticity is engineered to be small. A thermodynamic SPU explicitly programs the stochastic differential equation that governs the analog state and reads samples from its stationary distribution.[^11][^16]
Thermodynamic computing also differs from quantum computing in a way that matters for near-term commercialization. Both quantum and thermodynamic devices are designed to sample from a target distribution that would be exponentially expensive to draw from with a digital pseudo-random-number generator. Quantum computers, however, require cryogenic isolation from thermal noise; thermodynamic computers use thermal noise as the source of their randomness and therefore operate at room temperature in conventional CMOS-compatible processes. The trade-off is that thermodynamic computing has no analogue of quantum entanglement and so cannot, in principle, accelerate the same complexity class that a fault-tolerant quantum computer would. The Normal team has consistently framed the SPU not as a quantum competitor but as a probabilistic accelerator for the slice of AI workloads that are dominated by Gaussian sampling, matrix inversion, Langevin dynamics, and diffusion processes.[^15][^16][^19]
Normal's research output reflects an unusual concentration of physicists, quantum-computing veterans, and probabilistic-ML specialists in a single hardware company.
As of early 2026 the company reports more than fifty employees across four offices: New York City (headquarters), San Francisco, London, and Copenhagen.[^2][^4]
CN101 is a physics-based ASIC implementing the Carnot architecture. Normal has publicly disclosed that the chip targets linear algebra and matrix operations alongside stochastic sampling via Lattice Random Walk, and that the company expects the architecture to deliver up to a 1,000x energy advantage on suitable AI and scientific workloads.[^6][^7][^8] The tape-out announcement on August 12, 2025 marked the transition from prototype-on-PCB to characterization-and-benchmarking of a manufactured die. As of the announcement, Normal has not publicly disclosed the foundry process node, die area, or wafer partner.[^6][^7]
The technical novelty of CN101, relative to the published SPU prototype, is that the inductive coupling that gave the eight-cell PCB its dynamics has been replaced with resistive coupling and on-chip noise generators that are far easier to scale in standard CMOS. The chip is designed to be programmed at a high level using a probabilistic-graphical-model abstraction so that an application engineer writes the joint distribution to be sampled rather than the analog dynamics directly.[^13][^20]
Normal has published a public roadmap that anchors on three Carnot-architecture chips:
| Chip | Target | Workloads | Expected timing |
|---|---|---|---|
| CN101 | Foundational tape-out, characterization | 8x8 to small/medium-d Gaussian sampling, matrix inversion, LRW | Taped out August 2025 |
| CN201 | Production silicon for diffusion models | High-resolution image and scientific diffusion models | 2026 |
| CN301 | Advanced video diffusion models | Large-context video generation; long-horizon Bayesian inference | Late 2027 / early 2028 |
The roadmap is consistent with Normal's broader thesis that thermodynamic substrates are best suited to workloads dominated by Langevin-like dynamics, which is precisely the inner loop of diffusion models.[^7][^8][^21]
In parallel with the Carnot hardware, Normal has built a commercial AI-native electronic-design-automation product, marketed as Normal EDA. The platform applies large-model reasoning to register-transfer-level (RTL) design, verification, and ASIC implementation tasks, with the explicit ambition of compressing custom-silicon design cycles by roughly 2x in the near term and an order of magnitude over the medium term. According to CEO Faris Sbahi, by the time of the March 2026 funding announcement, Normal EDA was being used by more than half of the top ten semiconductor companies by revenue.[^4][^5][^9] Internally, the company uses Normal EDA to design its own Carnot chips, treating the EDA product as both a revenue line and a development accelerator for the rest of the business.[^4]
The company's published software stack abstracts the SPU as a primitive in the same way a GPU exposes matrix multiplication. The two primitives that CN101 exposes (linear algebra and LRW sampling) compose into a number of higher-level AI workloads:
For edge ai applications the company has hinted at lower-power variants of the architecture suitable for on-device probabilistic inference, but as of mid-2026 has not publicly released a dedicated edge product.[^7]
Disclosed rounds, in chronological order:
| Date | Round | Amount | Lead investors | Notable co-investors |
|---|---|---|---|---|
| June 2023 | Seed | $8.5 million | Celesta Capital; First Spark Ventures | Micron Ventures and others[^17][^18] |
| October 2024 | UK ARIA Scaling Compute grant | Selected (programme size £50 million across 12 teams) | UK Advanced Research and Invention Agency | n/a (non-dilutive)[^3] |
| Early 2025 | Seed extension / strategic | $35 million (cumulative) | First Spark Ventures (Eric Schmidt) | Celesta Capital, Drive Capital, ARIA, Micron Ventures, Samsung Next, Intel Ignite, National Security Innovation Network[^5] |
| March 2026 | Strategic | $50 million | Samsung Catalyst Fund | Galvanize, Brevan Howard Macro Venture Fund, ArcTern Ventures, Celesta Capital, Drive Capital, First Spark Ventures, Micron Ventures[^4][^9] |
Cumulative disclosed funding, as of March 2026, exceeds $85 million.[^4] The company has not publicly disclosed valuation in any of its rounds.
Normal Computing has an unusually visible research footprint for a hardware startup, with most key contributions posted to arXiv and several published in peer-reviewed venues. A representative selection includes:
Normal's most consequential partnerships, as publicly disclosed, fall into three groups.
The first is public-research funding. The October 2024 selection by UK ARIA for the Scaling Compute programme is the largest non-dilutive line of support; it co-funds Normal Computing UK's London office and gives the company access to a cohort of academic partners working on adjacent novel-architecture chips.[^3]
The second is silicon and supply-chain partners. The strategic round led by the Samsung Catalyst Fund in March 2026 explicitly linked Normal to Samsung's broader memory and foundry ecosystem, and the company's earlier rounds included Micron Ventures and Intel Ignite, both of which sit close to the production fabs Normal will need as it moves from CN101 to volume CN201 and CN301 silicon.[^4][^9]
The third is the Normal EDA customer base, which by March 2026 included more than half of the top ten semiconductor companies by revenue. The names of those customers have not been publicly disclosed, but Sbahi has confirmed both the count and the strategic importance of the relationships, since the same large foundries and integrated device manufacturers are the natural future buyers of Carnot-architecture IP.[^4][^9]
By 2025 the AI hardware market was crowded with novel-architecture startups arguing, in different ways, that the energy and economic curves of scaling nvidia GPU-based training had broken. Normal Computing fits into that landscape as the most prominent commercial vendor of thermodynamic computing, with one direct philosophical competitor and a wider set of indirect rivals.
The closest competitor is Extropic, a California-based startup founded the same year (2022) by ex-Google Quantum AI engineers Guillaume Verdon and Trevor McCourt. Extropic also pursues thermodynamic computing but emphasizes a "thermodynamic sampling unit" (TSU) architecture in which the chip natively produces samples from a programmable energy-based model, with claims of up to 10,000x energy savings for inference relative to GPUs.[^10][^11] Extropic emerged from stealth in March 2024 with $14.1 million in seed funding led by Kindred Ventures. The two companies differ in emphasis: Normal Computing presents its SPU as a general-purpose probabilistic accelerator for linear algebra, Bayesian inference, and diffusion sampling, anchored on continuous-variable dynamics; Extropic's TSU emphasizes a more directly generative interpretation in which the chip's stationary distribution is itself the trained model. Both companies share the underlying claim that programmable thermal noise can outperform programmable transistor switching for a useful and growing class of AI workloads.[^11]
The broader competitive table, drawn from public 2025-2026 disclosures, places Normal Computing among other novel-architecture chip startups.
| Company | Founded | Primary architecture | Target workloads |
|---|---|---|---|
| Normal Computing | 2022, New York | Thermodynamic (Carnot) / stochastic processing unit | Bayesian inference, diffusion sampling, linear algebra |
| Extropic | 2022, California | Thermodynamic sampling unit (TSU) | Generative sampling, energy-based models[^10][^11] |
| Lightmatter | 2017, Massachusetts | Silicon photonics interconnect and compute | Large-model training and inference |
| Mythic | 2012, Texas | Analog in-memory compute (flash-based) | Edge inference, vision models |
| Etched | 2022, California | Transformer-specific digital ASIC ("Sohu") | LLM inference |
| Rain AI | 2017, California | Mixed-signal neuromorphic / memristive AI accelerator | Brain-inspired training and inference |
| Tenstorrent | 2016, Toronto | RISC-V-based digital AI accelerator (Wormhole, Blackhole) | LLM and CV training and inference |
| cerebras | 2015, California | Wafer-scale engine | Large-model training |
| sambanova | 2017, California | Reconfigurable dataflow units | Enterprise foundation models |
| groq | 2016, California | Tensor Streaming Processor (TSP) | Low-latency LLM inference |
Within that landscape, Normal Computing's distinctive bet is the dual-product strategy. Most chip startups stake their viability on a single device generation; Normal monetizes Normal EDA against the wider semiconductor industry while iterating on Carnot silicon, and reuses each side to subsidize the other. The investor base, which by 2026 included Samsung, Micron, Intel, ARM, Brevan Howard, ArcTern, and Eric Schmidt's First Spark Ventures, is closer to a strategic semiconductor consortium than a typical venture syndicate, which both reflects and reinforces that strategy.[^4][^9]
Industry coverage in 2024 and 2025 generally treated the SPU prototype and the CN101 tape-out as credible early milestones rather than commercial products. Reports in Inside HPC, Data Center Dynamics, Tom's Hardware, TechRadar, and TechSpot focused on the company's published research, on its claims of up to 1,000x energy efficiency, and on the open questions about how those claims will hold up under independent benchmarking once CN101 silicon is in customer hands.[^7][^8][^27][^28] Normal itself has been comparatively conservative in framing CN101 as a "foundational" tape-out, with the production-grade chips (CN201 and CN301) still ahead of it.[^6][^7]
A second strand of commentary has noted that thermodynamic computing inherits a small but real legacy of skepticism from the broader probabilistic-computing community: similar claims have been made in the past for analog and stochastic accelerators that ultimately failed to scale beyond demonstration silicon. Normal's response, articulated most fully in its July 2025 landscape paper (arXiv:2507.10463) with collaborators at Cornell, Yale, Penn, and UC Santa Barbara, has been that the missing ingredient in earlier efforts was a complete software and EDA stack, and that AI-native chip design (the Normal EDA business) is precisely the missing link.[^25]