Share with your CTO
Sebastian Thrun (Google X, Waymo) and Adrian Aoun (Forward, Torch) are betting that foundation models can be retrained from generating code to generating physical engineering designs. Their startup Dulo raised $130 million in seed funding to build what they’re calling Engineering Superintelligence: AI that converts natural-language prompts into valid designs for cars, planes, and other hardware systems. Backers include Marc Benioff, John Doerr, DST Global, Construct Capital, and Lowercarbon Capital. The money goes toward GPU clusters and ML research hiring.
What this means for your business
The gap between software AI and hardware AI is a physics problem, not a compute problem. A model that hallucinates a function name costs you a failed unit test. A model that hallucinates a load-bearing specification costs you a recall, a liability, or a life. Dulo is explicitly targeting that gap, which means the first real question for any engineering-heavy enterprise isn’t whether to watch this space but whether your simulation and CAD data is clean enough to feed it.
The SimScale survey data from March 2026 makes the data readiness problem concrete: 80% of engineering organizations are running AI pilots, only 9% have a scaled program, and 74% cite data preparation as the primary barrier. That’s not a vendor maturity problem. That’s an internal data discipline problem that no amount of GPU spend by Dulo or its competitors will solve for you. Organizations using AI-enabled workflows already evaluated 56 design variants per program versus 17 for conventional teams. The productivity differential is real and compounding now, before anyone ships a mature foundation model for hardware.
The competitive dynamic worth watching: Prometheus raised $12 billion earlier this year for adjacent hardware AI work, dwarfing Dulo’s seed round by a factor of 90. Capital concentration at that scale typically means the winners won’t be foundation model builders but the enterprises that figure out how to integrate these models into existing PLM and simulation pipelines first. The signal worth watching is whether Dulo ships anything verifiable before the category gets absorbed by larger players who can bundle engineering AI into existing toolchains.
Concept deep-dive: Physics-constrained generation
Standard generative AI optimizes for statistical plausibility. Physical engineering design requires outputs that satisfy hard constraints: stress tolerances, fluid dynamics, thermal limits, regulatory safety margins. A model that ignores these produces designs that look correct but fail in the real world. Think of it as the difference between autocomplete and a licensed structural engineer. The challenge Dulo and its competitors face is embedding physical law as a first-class constraint during training, not as a post-hoc filter. For enterprise CTOs, this is why domain-specific training data, not general model scale, determines whether these tools ever reach production.
Based on reporting from Dulo Raises $130M for AI Engineering in San Francisco, originally published 2026-11-09 17:54:00.

