After the Prototype

Audrey Miller
Nov 2026

Most people still think of hardware as hard from start to finish. Over the past year, AI has made the first stage, going from an idea to a working prototype, fast and cheap. The second stage, turning that prototype into something you can legally sell and build by the thousand, is still slow and expensive for now.

A hardware founder told me recently that his last device took two weeks and about $10k to prototype thanks to AI but getting it through layout and compliance took the usual 18 months and $900k.

My AI-pilled bet is that, though harder, that second stage will get compressed next, leaving us with potentially boundless creation and new challenges in consumerism (and waste).


The prototype got easy

On 9 November, Anthropic showed three working gadgets that Claude designed from a short prompt: a speakerphone, a voice-assistant hub and a handheld virtual pet. The model drew the circuit diagram, laid out the circuit board, designed the case and wrote the software. Diode Computers, a manufacturer in Brooklyn, built all three from those files with no revision rounds.

Each model release makes this faster and cheaper but a working prototype is not a product.


But the product is still hard

Before you can sell an electronic device it needs certification, the official sign-off that it’s safe to sell. Every major market has its own compliance frameworks: FCC for the US, CE markings for EU, and IEC internationally. In the US, any device that has WiFi or Bluetooth, needs to satisfy FCC equipment authorization requirements, including rigorous radio-frequency testing, documentation of underlying components, and electromagnetic emissions testing, etc.

For example, every circuit leaks a little radio noise. So regulators require emissions testing, where an accredited lab measures that noise to check your device won’t disrupt the phones and radios nearby. Today, failure rates of these tests are at 70 to 90 percent for most devices’ first tests. Each attempt can cost tens of thousands of dollars, and the lab tells you pass or fail and little else.

Then comes system integration: making the board, case, battery and software work together in one box that a factory can build thousands of times. This is the process of turning a collection of working components, perhaps AI designed today, into a finished, manufacturable product. It’s one of the most underestimated parts of building consumer hardware. Electronics, mechanical systems, battery power, software, and manufacturing need to be integrated and tested for durability, yield, quality and many other parameters.

Code can be tested in seconds but a circuit board needs a days-long simulation, a factory run or a lab fee. AI can’t shortcut this yet because checking hardware is expensive. AI can’t physically flip the Samsung phone any faster in durability testing.


What happens when that gets easy?

None of this looks permanent to me. People are already working on predicting those results in software before anything gets built. Nominal, which builds testing software for hardware teams, raised at a $3B valuation in November. Quilter raised $25M for physics-trained circuit board layout, and Diode raised $11.4M, led by a16z, to design and manufacture boards under one roof. These are just a few.

If you are one of these people or building near this layer, I’d love to chat!

Making a sellable physical product will become something a small team can do in weeks. In that world, hardware gets cheap enough to build for niche audiences, from scientific instruments to the snow cone machines and the other useless side projects permanently living half-built across my desk.

For a generation, the know-how for taking a prototype to production has lived near the contract manufacturers and OEMs in Asia. This is one of the reasons we had conviction so early in Carl Pei building Nothing. He owned the know-how from his experience doing it all before, which let him run faster. If that know-how turns into software, building full-stack will be faster than ever.

Think about what that does to the things around us. Mass production only pays off at huge volume, so most of what we own was designed for cohorts, but not one person in particular. Take that constraint away and a product can be made for a few hundred people, a few dozen, or even for one: a keyboard shaped to one person’s hands, a sensor built for one lab’s experiment, a toy a parent designs with their kid.

Software has already been through this. When building got cheap, people started making small apps for themselves or their teams with no plan to scale them. Physical things will go the same way, with thousands of tiny brands serving audiences too small for a big manufacturer to notice. Before factories, most objects were made by someone who knew exactly who they were for. For the last 20 years, traveling to Venice or Strasbourg was watching a slow death of the small artisan shops, all shutting down and being replaced with multinational brands. I revel at a version of the future that reverses this thanks to technology.

Boundless creation has a cost too. When almost anything can be made cheaply, more will be, and cheap things get treated as disposable.

The world threw away 65M tonnes of electronics last year (and properly recycled less than a quarter of it... separate problem). Waste campaigners already talk about ‘fast tech’ the way the rest of us talk about fast fashion. If hardware gets as easy to produce as content, we could end up with the physical version of a feed full of slop. How do we prevent that version of reality?

And who repairs a device made by a three-person company in a run of 200, or answers for it when it fails?

When making gets easy, the hard part is deciding what deserves to exist.

If you’re working on any part of getting a prototype into production, let’s chat!