The New $8-Billion Bet on AI for MCAD
Opinion by Ralph Grabowski
With talk swirling of a LLM-based AI financial bubble collapse, taking place as soon as perhaps this year [source, source], investors are looking into different kinds of AI. “These [LLM] systems can never truly understand the world, because they can’t move through it,” says Patrick Vanbrabandt of Carya [source].
Last year, two companies announced they are spending eight billion dollars to model reality with World Model AI. This is a form of AI trained to understand how the real world works, such as through physics in 3D spaces.
World Model AI is best known for training self-driving cars, which need to be taught what can be driven on, and what must be driven around. Now World Model AI is expanding to include everything — so its financial supporters hope.
Project Prometheus
Six billion dollars went into a new CAD/CAM company, nicknamed Project Prometheus. The New York Times newspaper back in November was the first to reveal it, saying “The company is focusing on AI that will help in engineering and manufacturing in a number of fields, including computers, aerospace, and automobiles” [source]. The company is co-chaired and co-funded by Jeff Bezos, the former CEO of Amazon.
Sadly, that’s all we know, other than its tagline, “AI For the Physical Economy” [source]. The company remains silent. Still, we can piece Prometheus together from sources like employees hired and other areas in which Bezos is involved.
From their biographies on LinkedIn, we know that most of the 100 employees are in AI, such as autonomous intelligence and neural networks. A very few are from CAD-related fields, such as real-time simulation of industrial particulate flows. More are needed: co-founder William Guss made a public appeal on Twitter (X) for “Anyone in manufacturing and builds real things. Really trying to understand the space and see some factories :)” [source].
The other hint comes from Bezos’s involvement in manufacturing companies: Blue Origin’s reusable rockets and Slate’s low-cost electric trucks. My conclusion is that Prometheus wants to build an in-house CAD/CAM system assisted by a form of AI that understand the physical world to drive down development and manufacturing costs of products.
nVidia and Synopsis
Last December, Nvidia copied Bezos by buying $2 billion worth of Synopsis shares, the biggest electronic design firm — and new owner of simulation stalwart Ansys. Patrick Moorhead of Moore Insights says, “The partnership is a big tell for where AI likely goes next: engineering, simulation, and digital twins” [source].
A multi-year agreement places nVidia’s CUDA, Agentic AI, and Omniverse into Synopsis software. CUDA is the programming interface for nVidia’s graphics boards (GPUs), short for Compute Unified Device Architecture. GPUs are used for accelerating everything from games, to CAD, to crypto currency mining, and now for AI by running dozens or thousands of strands of programming code in parallel.
AI chatbots, like ChatGPT, handle a single prompt (question) at a time. Agentic AI works on multi-step problems. Agents are seen as a next stage in AI helping users, but come up against blockages, such as needing permission (like passwords) to access data and handling difficult-to-navigate Web sites. For this reason, the AI industry in December endorsed MCP [model context protocol] to “tell AI models which external tools, data sources, and workflows they’re able to access, then allow them to connect and perform tasks,” says AI reporter Hayden Field [source].
Omniverse is APIs [application programming interfaces] and SDKs [software development kits]. CAD vendors use it to illustrate real-time effects in massive models. For instance, SimScale deploys it to combine cityscapes with wind simulations generated by its CFD [computation fluid dynamics] software.
Omniverse displaying SimScale’s CFD-generated wind simulation
Hexagon ADAS
It turns out, however, that World Model AI is not all that new. Hexagon of Sweden has worked on it for close to a decade in its ADAS [advanced driver-assistance system] program. Hexagon’s reality capture software records physical structures along roads as 3D point clouds. Then its Virtual Test Drive software simulates vehicles driving on the roads, taking into account the 3D model, vehicle software and hardware, and the driver.
Hexagon’s reality capture recording physical structures
Hexagon tells me that it is committed to its virtual test drive software, updating it for the desktop and porting it to the cloud.
World Model AI In MCAD
MCAD systems, both old and new, boast of AI to stay relevant. Legacy MCAD vendors like Dassault Systemes and Autodesk state they’ve been working on AI for a decade already, but their reference point is generative design, which a decade ago they didn’t call “AI.”
AI in legacy MCAD systems tends to be chatbots that implement shortcuts: ask one question, and get one answer. A common scenario is asking the chatbot to render a scene with specific parameters; it’s a shortcut to tracking down commands -- like a search engine. Sometimes, there is no solution, because the chatbot hasn’t been programmed with the answer yet [source], a sign that we might not be dealing with artificial intelligence after all.
Autodesk says it has lots of AI in its Fusion design package, but further reading reveals that its AI tends to be automation -- replacing many steps with one, such as automating tool paths, automating drawings, and automating fastener replacements [source].
Paths for machining tools being automated in Fusion from Autodesk
The one firm I found furthest along is an architectural one, Snaptrude. With it, you enter a one-line prompt, such as “Design a 3-story office building with open workspace, meeting rooms, and cafeteria,” and then it spends ten minutes generating and optimizing an initial room layout that meets the specs of building codes -- happily chattering to you as it carries out the determinations.
Initial rooms layout generated through an AI prompt in Snaptrude
What Ralph Grabowski Thinks
The closest MCAD comes to World Model AI is with digital twins, where the CAD model is a replica of what will be built and has been built. It lets designers run simulations and optimizations on products like expresso machines and airplanes before building them. Adding AI could help identify problems designers might not have noticed [source]. The caution for us is that LLM-based AI can have a 30% failure rate by returning hallucinations -- information that it makes up [source].
Just like it is concerning that unexperienced engineers could access simulation software early in the design process, the same caution ought to apply to blackbox-AI assisting designers [source].
[This article first appeared in Design Engineering magazine and is reprinted with permission.]