The semiconductor industry has a problem that more power won't solve. Normal Computing is betting it can solve it with physics.
The New York-based startup announced $50 million in strategic funding on March 25, led by Samsung Catalyst Fund, bringing its total raised to over $85 million PR Newswire. The round includes Galvanize, Brevan Howard Macro Venture Fund, ArcTern Ventures, and existing investors Celesta Capital, Drive Capital, Eric Schmidt's First Spark Ventures, and Micron Ventures.
What makes the funding interesting is the problem Normal is trying to solve. CEO Faris Sbahi laid it out plainly in an interview with Fortune: "Data centers are expected to hit an energy wall around 2030. Most of the strategy now is to find new ways to acquire more energy — but our position is to solve the problem in terms of the hardware that we're using" Fortune. That framing — that the industry's dominant response to the energy crunch is the wrong answer — is the core bet the company is making.
The angle is worth sitting with. The dominant narrative in AI infrastructure right now is that power is the constraint and new power sources are the solution. Nuclear microreactors, offshore wind farms, PPAs for every gigawatt in sight. Normal's argument is that you cannot out-build the inefficiency. The thermodynamic approach is a different bet: instead of finding more electricity, change the computing architecture so you need less.
Normal has two products. The first is Normal EDA, an AI platform for semiconductor design that uses auto-formalization — combining large language models with formal logic — to compress chip design timelines. The company claims it can accelerate custom silicon to market by 2x today, with a longer-term roadmap toward 1000x efficiency gains through its Carnot hardware program PR Newswire. More than half of the top 10 semiconductor companies by revenue are already using the software platform, according to Sbahi Fortune. That customer list is the evidence the approach is not purely speculative.
The second product is the Carnot hardware program itself. In August 2025, Normal completed the tape-out of CN101 — which it calls the world's first thermodynamic computing chip, targeting multi-modal diffusion generative AI model inference PR Newswire. The tape-out was confirmed in a separate PR Newswire release last August and covered by Tom's Hardware and Data Center Dynamics. The CN101 is not a general-purpose processor. It is an ASIC designed to handle specific workloads: linear algebra, matrix operations, and stochastic sampling Tom's Hardware. The target applications are scientific simulations, optimization, and Bayesian inference.
The thermodynamic computing concept is genuinely different from how conventional chips work. Traditional GPUs — including NVIDIA's Blackwell architecture and AMD's CDNA-based accelerators — process computations by suppressing or managing the inherent randomness in physical systems. They spend energy fighting noise. Thermodynamic computing works with those dynamics: fluctuations, dissipation, and stochasticity are not problems to solve but resources to use PR Newswire. The chip uses the natural behavior of physical systems to compute more efficiently for certain classes of problems. That is not marketing language — it describes a real architectural distinction that has been explored in academic literature for years. Normal is one of the few attempts to build production silicon around it.
The hardware program has external validation. The Advanced Research + Invention Agency (ARIA), the UK government's high-risk, high-reward research funder, has supported the Carnot program PR Newswire. Dr. Suraj Bramhavar, ARIA's programme director for the Scaling Compute programme, said in the press release: "Normal's team has taken a fundamentally unconventional approach and delivered working silicon in CN101. That is an exceptionally rare outcome for work this ambitious."
The competitive field is not empty. Unconventional AI, led by Naveen Rao, former head of AI at Databricks, raised a $475 million seed round in December 2025 led by Andreessen Horowitz and Lightspeed Ventures Fortune. Extropic is developing probabilistic AI chips based on a different technical approach. Both are betting that the post-GPU computing era will require different silicon. Normal's distinction is that it is pursuing both the software layer and the hardware layer simultaneously — using its own EDA tools to design its own chips.
There is a reason Samsung is leading this round. Samsung Catalyst Fund invests strategically in semiconductor-adjacent technology that could affect Samsung's own manufacturing and chip business. Micron is also an existing investor — memory and AI infrastructure are tightly coupled. The investor list signals that major semiconductor manufacturers see Normal as having a real shot at the efficiency problem, not just another AI chip startup promising faster training.
Whether thermodynamic computing delivers on the 1000x efficiency roadmap — versus the 2x near-term software acceleration, which is more credible given existing customer traction — is genuinely open. The CN101 tape-out is a real milestone. Getting from tape-out to a production chip that data centers actually buy is a different challenge. The energy wall around 2030 is also a projection, not a certainty; it depends on AI training scale continuing to grow at current rates, which is itself being questioned as inference economics begin to dominate.
What is not in doubt is that the problem is real. AI inference at scale is an energy problem, and the industry knows it. The question is whether the answer is more power infrastructure or better silicon. Normal is betting on both — and has the Samsung money and the customer traction to make the bet worth watching.
