A few years ago I spent some time implementing (by hand) some algorithms for geometry operations as an academic exercise, the primary result of which was a great admiration for people who build CAD software. I ran into the former CTO of Adam at a WeWork recently, which got me thinking some more about CAD. Adam was in the YC W25 batch and built a tool for doing text-to-CAD. It was interesting to talk to him about his learnings from working in the space.

Background

When I was working on K-Scale, I spent a decent amount of time in CAD, specifically Onshape. We chose Onshape initially because it was free for public projects and because it was cloud-native, which made it the best choice since we were using Macs and using something else for CAD would have broken my flow. I am probably not representative of most people using CAD tools for mechanical engineering, but it did give me some insights into some common pain points in CAD software. Specifically:

1. Onshape gets bogged down with big assemblies because the constraint solver gets overloaded. Onshape uses a custom version of the same Parasolid kernel that SolidWorks uses, but it runs in cloud rather than locally.
2. There's a huge amount of repetitive work when doing CAD that feels ripe to be accelerated. However, a full text-to-CAD tool would probably not be very useful, for the same reason that Cursor is a better product than Devin.
3. CAD software pricing has room to be disrupted. It mostly relies on upselling enterprise customers for legacy software. I think it should be pretty easy for someone reasonably smart to build a native Onshape alternative.

I think there's a market for an Onshape alternative that can do something more narrowly-scoped more effectively than Onshape currently does. For example, a lot of hobbyists today do stuff with sheet metal, thanks to SendCutSend making sheet metal prototyping cheaper and faster, but Onshape's sheet metal tools aren't that good. That type of thing might provide a good toe-hold to kickstart word-of-mouth adoption, if the core product is good enough.

It's a bit different from building a zero-shot text-to-CAD program since competing with Onshape would require doing some pretty hardcore engineering. CAD software is pretty tricky to build effectively and performantly. Out-performing Onshape would require writing some GPU-accelerated Parasolid alternative, which would be pretty hardcore engineering. Zoo is kind of trying to do this, but I don't think Zoo has really nailed the web experience yet. I'd probably try to write some constraint solving kernels in CUDA-accelerated Python, maybe with CuPy or even Triton kernels on top of PyTorch. Given that it's all running on the backend, it wouldn't matter too much to have a bunch of package dependencies. It's easier to run cloud GPUs today than when Onshape was originally getting architected. It would also be nice to do GPU-accelerated FEA on the backend, since this experience is currently pretty slow and annoying.

Eventually we should shoot for some Figma-style go-to-market strategy to unseat existing incumbents. The way that I ended up choosing Onshape instead of SolidWorks or Fusion actually resembles Figma's early adoption patterns pretty closely, so I think some of the guiding principles transfer quite well:

• Bottom-up, not top-down
• Collaboration-first, not feature-count-first
• Self-serve + viral loops, not channel sales
• Instant onboarding, not multi-week trials and reseller demos

Sprinkling on AI

Assuming it's possible to deliver some better-than-Onshape performance for something like sheet metal design, the next step would be to figure out how to use an LLM like Qwen to automate certain parts of the user flow. This kind of user augmentation experience should be possible to build into the IDE nicely, with a Greptile-like RL feedback loop to make the suggestion model better.

There's good reason to believe that existing frontier models aren't that good for CAD automation, for a similar reason to why existing models suffer for generating SVGs. Infra layers like Browser Use make it easier for AI models to interface with some tools, but CAD software isn't currently very amenable. I spent a lot of time messing with the Onshape API at K-Scale. There have been projects like building an Onshape MCP server but they haven't been very useful, for the basic reason that Onshape is not an AI-native IDE - the representation that you feed to the model just isn't that great, and the outputs that the model gives you are basically useless.

The dream scenario is that CAD can provide an interesting and useful environment for teaching AI models something about 3D relationships. Maybe this is how we will end up nailing SVG generation.

Existing Players

Cloud CAD:

• Onshape - The main player in cloud CAD, owned by PTC. Still CPU-based, which is why it gets slow with large assemblies. They're not going to build GPU acceleration because it would require rewriting their entire rendering engine, and they probably don't want to cannibalize SolidWorks.
• Fusion 360 - Autodesk's cloud-connected CAD, but it's still primarily desktop-based with some cloud features. Not truly cloud-native like Onshape.

Modern CAD UX:

• Shapr3D - iPad-based CAD that's actually quite good. GPU-accelerated on the device, focused on industrial design rather than engineering. Not cloud-based.
• Plasticity - Modern desktop CAD with great UX, focused on subdivision surface modeling for game dev and 3D art. Not parametric engineering CAD.

AI-first CAD:

• Zoo - This is actually pretty interesting. They built their own geometry engine (not using Parasolid or ACIS), their own programming language (KCL), and their own ML model that generates parametric B-rep models (not just meshes). They have both a GUI and a code view. The interesting part is they can fine-tune their model on your company's CAD files to learn your design patterns. But their frontend is clearly much worse than Onshape, which makes me think they're not really trying to compete on the interactive design experience - more focused on programmatic/ML workflows.
• Adam - YC W25, doing text-to-parametric-CAD with a conversational interface. You describe what you want, it generates a parametric model with sliders you can adjust. They're planning to build a deeper co-pilot that integrates into existing CAD tools. Founders have AI research backgrounds (Adept) plus mechanical engineering. This seems more focused on rapid prototyping and ideation than replacing professional CAD.
• Leo AI - Building a "large mechanical model" trained on engineering data that can suggest parts and generate assemblies. Don't know much about them yet because their app crashed when I tried to use it.
• Various text-to-CAD tools - There are a bunch of startups (OpenArt, AdamCAD, Vondy, DraftAid) doing text-to-2D-CAD, but these are more like design assistants than full CAD tools.

Enterprise AI CAD:

• WSCAD ELECTRIX AI - AI-powered electrical CAD, claims to reduce design time by 99%. Very domain-specific.
• Siemens NX with AI - Adding ML and NLP to an existing enterprise CAD tool. Probably slow to ship because it's enterprise software.
• Ansys with AI - Simulation-focused, not really core CAD.

The interesting thing is that nobody is doing the full stack: GPU-accelerated cloud CAD + AI features built from the ground up. Everyone is either doing cloud without GPU acceleration (Onshape), GPU acceleration without cloud (Shapr3D), or AI without good infrastructure (the text-to-CAD startups). The latter might end up backfilling the infrastructure, but I think real PMF will be basically bottlenecked until this is built, and there's room to compete here on technical capabilities.