Pinned Tweet
Manufacturing is due for a step-change.
I can't share what I worked on in 2025, other than building things that fly very fast.
Looking forward - it's time to build the machine that builds the machine
open.substack.com/pub/willqy…
If only there was a company which could scale hardware manufacturing with no upfront cost
Defense-tech companies face the hardest chicken or egg fundraising challenge in today's environment
The @DoWCTO wants companies to prove they can deliver at scale which conflicts with traditional early-stage investing:
> A company can only deliver at scale if they have capacity
> Company only attracts capacity if they've raised enough capital to invest in capacity
> Company can only raise enough capital if strong demand signal exists
> But there's no demand signal if you can't prove delivery of enough supply
I think most venture investors don't have the stomach to continuously capitalize a company until all these signals are clearly visible.
1
5
278
Me when I build a rocket and I realise touching faces in CAD isn't a pressure seal irl (true story) (they thought they could 3d print a pressure vessel)
this is exactly why assembly CAD is not verifiable and why LLM CAD will plateau after single parts
no amount of FEA, CFD or XYZ will get you to verification. everyone has a plan (analysis) until they get punched in the face (actually building the thing)
9
405
Will Yu retweeted
I've backed 9 true robotics companies (not including e.g. non-robotics hardware businesses) so far this year, 6 of which are in Europe (CH, DE, UK) vs 2 in the US (SoCal, NYC).
I routinely take crap for this opinion, but imo the talent starting companies in European robotics is just much stronger than what we have in the US atm. At least I'm putting money where my mouth is!
Will Yu retweeted
Low-volume, high-mix manufacturing presents 3 fundamental challenges for automation:
- Design: Products aren’t designed for automated assembly.
- Deployment: Automated assembly lines are bespoke to individual products, with little skill transfer to new products.
- Flexibility: Adapting lines to changes in product design, supply variance, etc. is incredibly difficult.
We’re fixing this. More on it soon.
A top-level @anduriltech exec shared at dinner tonight that they use little to no automation in their manufacturing facilities because “automation today is incapable of adapting to low-volume and high complexity production lines”
There’s still such a gap between the automation dream and what is technically possible today.
Within a year, AI will be designing everything you can imagine. But to produce them at scale is a different challenge
2026 is the year AI learned to engineer the physical world.
Within a year, I believe AI will design robots, satellites, and vehicles.
After 7 years building Flux, here's why I think hardware engineering and the CAD industry are about to change forever.
1
66
Know your product and know your robot, and you shall build one thousand factories
There’s an enormous amount of capital going into robotics companies that promise to fix American manufacturing by deploying robots into existing factories.
I think we’re solving the wrong problem.
When a new component comes in, someone has to figure out how to manufacture it. What machines to use, how to hold it, what tools are needed, how to inspect it, and what to change when things go wrong.
After all that work, a robot can load the part into the machine and take it out when it’s finished.
We’ve spent decades making machines faster and automating the physical work, but the thinking behind manufacturing remains almost entirely human.
And now we’re spending enormous amounts of money automating some of the cheapest tasks in that process.
Robotics has an important role to play, but deploying robots into factories designed decades ago won’t fundamentally change how manufacturing works.
We should be building machines that can figure out what to do in the first place.
1
135
I built something like this back when I was working on subsonic missiles. Immensely useful to relate requirements -> engineering data -> CAD parameters to sweep the entire design space / zero in on the correct design
E.g. how does your wing design affect your min launch speed?
Will Yu retweeted
watching LLMs replicate UI’s of existing CAD systems and calling them “AI-native CAD” is like Henry Ford watching people build these and call them “Engine-powered horse’s”
the reason why digital twins are bullshit is simple:
alignment drift between records and the real world.
But it doesn't need to be this way.
1/
Not just for Lego, but CAD programs, Process Simulate, Delmia - you can already design the factory in digital, just like you design parts. But it takes time and the sim-to-real is hard.
8/
1
1
49