Cheap boards, expensive everything else

In October I asked on X what will still protect a consumer electronics product once hardware starts to behave like software, with prototypes so cheap that anyone can build what they want. Twelve people voted. Half picked manufacturing know-how, 41.7% picked real-world deployment, 8.3% picked certification costs, and nobody picked branding.

Twelve votes prove nothing on their own. They do match what the last three months of building AnchorCam Stereo taught us, so here is the argument with our own numbers.

In August, a board cost weeks of other people’s time #

We announced AnchorCam Stereo at the Founders, Inc. demo day on 8 August: a stereo camera with an IMU, a microphone, Wi-Fi and storage on one board, made for capturing robot-learning data. Two days later I was asking on X whether any factory in Shenzhen could build a double-sided six-layer board in three or four days, fabrication and assembly included.

On 24 August I posted that we were delaying shipping. Customers had asked for changes, the distance between the two cameras grew from 50 mm to 70 mm, and that meant redesigning the board from scratch. The engineers we were hiring in Shenzhen told us to go slow, because a mistake in a hardware design cannot be debugged once it reaches mass production.

That was good advice for the way boards were made. A layout pass took skilled people and weeks of calendar time, so every pass had to be right. Then we lost the ability to make passes at all: when my co-founder left, we had no hardware engineer on the team. I looked in SF, Shenzhen and Seoul for a firm that could deliver a production-quality egocentric camera to our specification and did not find one. For a while all I could do was assemble other companies’ off-the-shelf boards.

The layout pass got cheap #

So we built our own tools. Zap! is our PCB router. On a public set of 25 benchmark boards, given 10 minutes and one CPU core per board, it fully routes 14. Freerouting routes 11 and tscircuit routes 10 under the same limits. Version 1.1 cut unconnected connections by 3.7 times across 50 boards it had never seen. When Zap! finishes, AI agents usually clean up what is left.

The current AnchorCam Stereo board, which we call P28, was placed and routed by our own AI tools. I posted the first fully AI-routed version on 7 October. JLCPCB assembled it, and assembly finished on 23 October. The board boots Linux and records both cameras. On the first P28 board, every IMU sample lands within 14 µs of the camera trigger, with a median of 3 µs. Across ten recordings, about 37 minutes in total, we lost one camera frame out of about 65,000.

The enclosure went through the same kind of loop. Three prototype shells in a row weighed 111 g, 99 g and 89 g.

The board did not become easy. What changed is the price of being wrong: a layout pass now costs hours of compute where it used to cost weeks of a contractor’s time. “Go slow” was the right rule when each attempt was expensive. With cheap attempts, we make more of them and test every one on a real board.

Where the time goes now #

The time went into the two answers most of the poll picked.

Manufacturing know-how first. On 25 August the global-shutter sensor we had designed around went from 120 yuan to 300 yuan in one week. We switched to the AR0234, which has a higher resolution at the same 3 µm pixel size, and planned to stock 1,000 sensors that day for 500 devices. A router cannot tell you which part to switch to or where real stock exists. Debugging is similar. In September our test board kept boot-looping, dropping into its recovery mode and falling back to USB 1.0 speed. We chased it as a software bug for days before finding that the cause was on the board. Cheap layout does not shorten that kind of search, because it happens with a unit on the bench.

Deployment is the other one. Before AnchorCam Stereo we built a camera that signs each frame to prove the footage is real, aimed at deepfake fraud in remote job interviews. It did not sell. The person who would hold the camera, the candidate, was not the person with the problem, and the hiring company had no way to make candidates use it. A cheaper prototype would not have changed that. We learned it only by putting the device in front of the people it was for.

What this means for a small hardware company #

If our board is cheap for us to iterate, it is also cheap for someone else to copy. What stays scarce is the work around the board: knowing the supply chain well enough to survive a price spike, testing on real units, and getting devices worn by real people.

For us that last part matters most, because of what the devices produce. In August I wrote that LLMs scaled on free data, while physical AI has to pay for every hour of synchronized multimodal data. That is still the bottleneck for robot learning, and cheap board iteration does not remove it. What it does is let a team of our size put more capture devices into the field sooner and spend its hours on the problems that only show up there.

Our loop for now is short: route with Zap!, build at JLCPCB, test the boards on the bench, change the design and order again. We spend most of our time on the parts of the job that this loop leaves untouched.