Vibe coding wrote your software. Vibe manufacturing is now ordering your steel.
A real-estate developer in East Lansing needed a small steel piece to secure a camper to his truck. He described it out loud — the part, the angle it had to sit at — to GPT-6 Astra in voice mode. Five minutes later the model had a prototype. It exported the design as a DXF, opened the website of a sheet-metal fabricator called SendCutSend, filled in the order form and paid the 156 dollars with the payment details he had stored. Ten days later the part was on his doorstep. "In about 10 minutes," Jeff Deehan told Business Insider, "I spoke a part out of my mind into existence, without touching my phone."
That is the pitch, and Business Insider collected five versions of it in one piece this week: a robot arm, a 3D-printed brain from an MRI scan, a children's toy, an answering box for the grandparents. The builders call it vibe manufacturing — the hardware sequel to vibe coding — and the pattern repeats: describe the object, let the model emit the geometry, send the file to a fabricator or a 3D printer, iterate until it fits.
What the stack actually looks like
The interesting part is not the chatbot, it is the file. Manufacturing does not take vibes: SendCutSend's own FAQ is blunt about needing a proper CAD file — DWG, DXF, STP or STEP — before anything gets cut. So the trend is really about models that produce manufacturable geometry.
- Amy Zhou, an ex-Apple iPhone case designer now at the rendering startup Vizcom, and Jake Fitzgerald are building an open-source tool for exactly that. They call it text-to-CAD; it already tops 23,000 stars on GitHub.
- Commercial versions exist too: Zoo's Text-to-CAD turns a description into real CAD through an API, and the whole subject already has its own playbook site with thirteen steps — "have a frontier model write the design as code".
- The playbook's advice is the most grown-up thing in this story: have the model emit OpenSCAD or CadQuery instead of an opaque model, give every dimension a name, keep the design in a git repository so each iteration is a commit, state the manufacturing process up front, and — the line worth framing — ask the model to list every assumption it made, then challenge each one.
Code-first CAD is what makes the loop survivable. A parametric file can be diffed, reviewed, re-run with new measurements. A pretty 3D render cannot.
What it costs before it works
Fitzgerald's robot arm — sheet metal, about 200 dollars, 0.6 metres of reach, a motor and a camera, built as an "edutainment" product to learn robotics before the industry booms — has a prehistory. A few months earlier he watched the AI stack "compounding failures": bad designs built on bad designs, screw holes arriving loose. He spent around 2,000 dollars ordering pieces that did not work.
David Plakon, in Austin, printed the case for his children's answering box about fifteen times. Some pieces did not fit together at first. "AI's good but not perfect," he said. Arthur Prince, outside Paris, modelled his daughter's toy in a week — months of work by his own CAD skills — but only after running one chatbot past the other: GPT-6 Astra did the casing and the wiring, Claude was asked to criticise it, and he ordered ten units from JLCPCB for about 500 euros once the two models agreed.
So the honest summary is not "anyone can make anything". It is: the first draft is nearly free, and the last twenty percent still costs money, prints and evenings. Complexity scales badly — nobody is vibe manufacturing a car at home.
Where the scepticism is correct
Two pieces of research land right in the middle of this.
CADEngBench (arXiv, Mohamed Bin Zayed University of Artificial Intelligence) exists because image-comparison benchmarks are useless here: "An AI-generated part can look correct even when its program fails" — a named dimension controlling the wrong feature, an edit that quietly damages a mounting hole, a simulated response that differs from the reference. Their point is structural: CAD is not a picture, it is geometry plus an executable engineering model. Testing that a shape looks right establishes almost nothing.
MIT's state-of-practice study (over 30 interviews across large firms, SMEs, AI developers and CAD/CAM vendors, March 2026) found adoption is constrained less by model capability than by fragmented and machine-unfriendly data, security and regulatory requirements, and legacy toolchains with no usable API. Reliability, verification and auditability are what gate deployment — which is why the human-in-the-loop review stays.
Meanwhile the professional side is already automating its own intake: Xometry's AI-powered process recommender reads a customer's CAD file and suggests the process. The humans on the other end of the file are the ones asking the uncomfortable questions — in engineering forums, the recurring joke is "are we cooked?".
The labour question, with numbers
One of the builders is remarkably direct about it. Sourish Sharma, 20, from London, turned his own MRI data into a 3D-printed brain with a single prompt, after an open-source tool did not work well by hand. His verdict: "AI literally does replace CAD engineering completely. It's kind of unfortunate if you're a CAD engineer. Fortunately, I'm not."
Autodesk's 2026 AI Jobs Report complicates the simple version of that. AI hiring in the design-and-make industries more than doubled in two years (+147%, and another +33% in the last year), but the fastest-growing titles are not the ones replacing engineers — they are creative: AI UX Designer at +145%, AI Creative Technologist at +123%. The report's own framing is that the skills gap is the bottleneck, not the headcount.
What to take from it
Vibe manufacturing is real, and it starts where the stakes are low — toys, jewellery, a bracket, a box for the grandparents. That is not a small thing: it is the same path vibe coding took, from silly demos to production. What separates the two is that a buggy web app gets a patch, and a badly dimensioned steel part gets a refund, a bent machine, or worse. The playbook that is emerging — code-first, named parameters, versioned designs, assumptions listed and challenged — is the sound engineering answer to that, and it is the part worth stealing whether or not you ever ask a chatbot to buy steel for you.
Sources: https://www.businessinsider.com/vibe-manufacturing-ai-prototypes-cad-coding-2026-11 https://vibe-manufacturing.com/playbook/design-as-code https://arxiv.org/abs/2608.09296 https://decode.mit.edu/assets/papers/2026_Edwards_Agentic_AI_in_Engineering_and_Manufacturing.pdf https://adsknews.autodesk.com/en/news/2026-ai-jobs-report https://sendcutsend.com/faq/what-if-i-dont-have-a-cad-file/ https://investors.xometry.com/news-releases/news-release-details/xometry-updates-process-recommender-its-ai-powered-instant https://zoo.dev/text-to-cad
Speak the part, then check the fits.
