JuliaHub has closed a $65 million Series B and launched Dyad 3.0, an agentic AI platform for industrial digital twins that uses physics-based simulation and autonomous agents to compress engineering design cycles from months to days.
Most of the conversation around enterprise AI has stayed safely on the software side: writing code, summarizing documents, automating workflows that live entirely inside a computer. JuliaHub is working on something considerably more demanding. The company, which builds scientific computing infrastructure on top of the Julia programming language, announced yesterday that it has raised $65 million in a Series B round led by Dorilton Capital, with General Catalyst, AE Ventures, and former Snowflake CEO Bob Muglia all participating. Alongside the funding, JuliaHub launched Dyad 3.0, an agentic AI platform designed to help engineers design, simulate, and generate control code for physical systems. Satellites, heat pumps, semiconductors, industrial machinery. The kind of engineering that has historically required teams of specialists and timelines measured in quarters.
The funding round's composition is worth noting. Dorilton Capital leading the round signals serious industrial backing, and Bob Muglia's personal participation carries its own weight. Muglia ran Snowflake through some of its most consequential growth years before its landmark IPO in 2020, and his involvement in a scientific computing company pivoting toward agentic AI for physical systems suggests he sees a genuine enterprise category forming here, not just a research curiosity.
Dyad 3.0 combines three capabilities that have historically lived in separate tools and separate teams. The first is physics-based simulation, the kind of high-fidelity modeling that engineers use to predict how a system will behave under real-world conditions before any physical prototype exists. The second is scientific machine learning, which layers data-driven models on top of those physics simulations to improve accuracy and handle complexity that pure equation-based modeling struggles with. The third is the agentic layer: autonomous AI agents that can move through design iterations, test configurations, and generate the control code that tells a physical system how to behave, without requiring a human engineer to supervise every step.
The practical implication is that an engineer working on, say, a new heat pump configuration could describe the design constraints and performance targets, then let Dyad's agents run through simulation cycles, identify the most promising design paths, and produce working control code for the preferred configuration. Work that previously required coordinating across mechanical engineering, controls engineering, and software teams over several months becomes something that can happen in a compressed timeline with a much smaller team in the loop.
JuliaHub has been building toward this for several years. The Julia language itself was designed for high-performance scientific computing, and the company has cultivated a user base of engineers and researchers who were already running serious computational workloads. Dyad 3.0 is not a product built on top of a general-purpose language model with some physics vocabulary bolted on. It is built on infrastructure that was purpose-designed for the mathematical demands of physical simulation, which is a meaningful technical distinction when the alternative is asking a system trained primarily on text to reason reliably about thermodynamics or orbital mechanics.
Why physical AI is becoming an enterprise priority
The category JuliaHub is staking out has been quietly building momentum. Nvidia has made no secret of its ambitions in physical AI and industrial digital twins through its Omniverse platform. Siemens and Dassault Systemes have been integrating machine learning into their simulation tools for years. What is new is the agentic layer: AI that does not just assist an engineer but can autonomously navigate a design space, generate solutions, and iterate without continuous human direction. That capability changes the economics of engineering R&D in ways that are still being worked out.
For industrial companies, the appeal is straightforward. Engineering talent is expensive, specialized, and in short supply across most of the sectors Dyad 3.0 targets. If an agentic system can handle the iterative, computationally intensive work of design exploration and controls generation, senior engineers can focus on judgment calls that genuinely require human expertise: setting constraints, evaluating tradeoffs, validating outputs against real-world experience. That is a different value proposition than automating a knowledge worker's email. It is closer to giving a small engineering team the throughput of a much larger one.
The $65 million gives JuliaHub the runway to build out the customer base and the platform simultaneously, which is the right sequence. Enterprise software for physical engineering has a long sales cycle and requires deep integration with existing simulation and product lifecycle management tools. Getting that right takes time and sustained investment. With General Catalyst in the round alongside the more industrially focused capital from Dorilton and AE Ventures, JuliaHub has backers who understand both the software scaling story and the industrial adoption reality.
The question now is how quickly engineers at large industrial companies get comfortable delegating design iteration to an autonomous agent. That trust is built one validated output at a time. If Dyad 3.0 delivers consistently on the compression it promises, the case for physical AI as a serious enterprise category will start writing itself.
Also read: AMD's compact Halo Box could give small AI teams a serious alternative to renting cloud GPUs • KAIST Researchers Found That Teaching AI to Embrace Chaos Before Training Makes It Dramatically More Honest About What It Doesn't Know • Chinese Courts Just Ruled That AI Automation Is Not a Legal Reason to Fire Someone and Every HR Team Needs to Read This
