If AI makes designing physical things cheap, what should a person with software skills and no factory do? A short, hedged reading of the research: why to try, how to start small, what to avoid, and what would change the answer.
If AI makes designing physical things cheap, the question worth asking is what a person with software skills and no factory should do about it. This is a short reading of a longer, sourced research document. It is a reasoned guess and not a result, and the research under it is thin in places.
"Vibe manufacturing" is used for different things. An industry association uses it for AI plus additive manufacturing inside existing companies. Here it means one person, with agents, getting a physical thing made. I found nobody else using it that way yet.
Why bother
The premise is easy to like: design gets cheap, and making becomes the constraint. The evidence is mixed. On benchmarks that score past whether the code runs, AI-generated boards and CAD parts pass far less often than they run. The best full pass rate on one hardware benchmark is 8.15%. What looks cheap is drafting. What still costs is checking.
That is a reason to get involved anyway. Checking is a software-shaped problem. Design-rule checks, simulations and vendor manufacturability checks are deterministic, and the stronger results in the research pair a model with one of them.
The second reason is that trying is cheap. A small circuit board is advertised from $2.00 for five, and assembly from an $8 setup fee. Those are one vendor's starting prices, not typical ones, but a first order costs little. No source measures what first orders teach, so the way to find out is to place one.
How
- Pick something small you would use yourself: a board or a bracket, not a product.
- Design it with an agent, in tools an agent can drive. Code-based CAD and KiCad scripting both have working open libraries.
- Put a check in the loop before you order: design-rule checks, simulation, the vendor's own manufacturability check. Write down what each one caught. In one documented attempt an agent nearly chose a timer chip that does not run on a 3 V coin cell, and the same write-up logged a via placed almost on a pad. That attempt had a person and a rule checker around the agent.
- Order the smallest batch the vendor allows and read the limits first: minimums, lead times and, if you are in the US, tariffs. One vendor's own FAQ gives an overall range of about 35% to 92.5% for US buyers of Chinese parts as of March 2026.
- Write down what went wrong. No source I found measures how often first-time customers fail, so your own log fills a gap.
What I would stay away from for now
Anything that needs capital. In one Kickstarter sample of successful design and technology projects, 24.9% delivered on time, and the founders named manufacturing problems, shipping and certification. Quirky and Shapeways, which carried making or a marketplace at scale, both went bankrupt. In earlier cost collapses, trouble clustered at making and delivery, not at design.
Software for factories and data tooling for robots are funded, but by large rounds: one factory company raised $1.37B in August 2026. I found no evidence of small teams with traction in either, which is not the same as there being none.
What would change this reading
- A study that takes AI-generated designs through fabrication and counts what worked.
- A small shop shown to reach several times its output at the same headcount. The claim is circulating and I found no case.
- Measured movement of office or software workers into factory or data-centre work. The best labour study I found sees no clear AI disruption yet, apart from workers aged 22-25 in exposed jobs.
The research document has the figures, the sources and a list of what I could not find. If you try a first order and it goes wrong in an interesting way, I would like to hear about it.