If you asked me last year how a person on a small budget should do computer vision on a physical thing, I would have given you the same answer everyone else gives. Get a Raspberry Pi 5. Put the AI HAT+ on it. That gets you thirteen or twenty-six trillion operations a second sitting on top of a hundred-dollar board, wired into the camera stack, with tutorials for everything.
It is a genuinely good answer. I have recommended it to people. It is still a good answer this morning.
What I did not know, the whole time I was giving that answer, is that the company making the chip on that HAT was taking emergency loans to stay alive.
What Actually Happened
Microchip Technology signed a definitive agreement to acquire Hailo. The announcement went out July 24th and turned up in the Edge AI and Vision Alliance's roundup on August 5th, which is where I saw it. Terms were not disclosed. The deal is expected to close by the end of September, subject to regulatory approval. Microchip noted it does not expect a material effect on its financial results.
Read that last sentence again, because it is the whole story compressed into corporate boilerplate. Not material.
Hailo is not a nobody. Founded in 2017. Hailo-8, Hailo-10, Hailo-15. Their silicon is in the Raspberry Pi AI Kit and the AI HAT+, which is to say it is in the hands of more small builders than almost any other AI accelerator on earth. Microchip's own release counts more than a hundred current customers and a developer community of more than ten thousand.
Here is the part the release does not mention, which the Israeli business press did:
| Item | Figure |
|---|---|
| Total raised since 2017 | about $340 million |
| Peak valuation | above $1 billion |
| Valuation going into the deal | under $500 million |
| Sale price | undisclosed; reported as a fraction of prior valuations |
| How it got there | liquidity crisis, emergency loans, major workforce cuts, a collapsed SPAC merger |
Delek Automotive, one of the backers that provided emergency financing, has said it will write off most of its investment.
So: a company with real silicon, real customers, and real design wins, in the hottest category in hardware, in a year when robotics startups raised something on the order of fifty billion dollars and a humanoid company with no revenue to speak of got marked at thirty-nine billion. And it sold for scrap.
Nobody Downstream Could See It
This is the part that bothers me, and it is not schadenfreude. It is a hole in how I evaluate parts.
Through all of that — the layoffs, the bridge loans, the dead SPAC — the product kept shipping. The HAT stayed in stock. The GitHub examples kept getting commits. The forum stayed busy. Somebody wrote a new tutorial. From where I sit, holding a board and a soldering iron, there was no signal at all. The distress was completely invisible from the outside right up until the day an acquisition notice appeared.
I have a spreadsheet for every part in the sump pump monitor and in the forty-device building deployment. Price. Availability. Power draw. Accuracy. Whether the SDK is tolerable to work with. Lead time.
There is no column for will this company exist in five years. There never has been. I would not know how to fill it in honestly if I added one.
That is a rule I can actually apply. A commodity sensor from a dead company is still a sensor. A proprietary compiler from a dead company is a brick with your product attached to it.
Why My Pump Monitor Didn't Notice
I have made this point before about Nvidia's pricing and I am going to make it again, because this week gave me a better version of the evidence.
My sump pump monitor has no AI accelerator in it. It has a fifteen-dollar smart plug watching the motor's current draw. When the pump runs, the current tells you how hard it is working, how long the cycle lasted, and how often it cycled. A motor with a failing bearing pulls more current before it pulls its last. That is the whole trick, and you can hold the math in your head.
It has been running for two years. It has never had a supply chain. It has never had a vendor. The smart plug has four interchangeable competitors at the same price, and if all of them vanished I could measure current with a clamp and a microcontroller for less money and more annoyance.
None of this makes me clever. I did not choose that architecture because I foresaw the edge-AI shakeout. I chose it because I was cheap and the problem was small. But the thing about being cheap on purpose is that it keeps paying out in ways you did not plan for, and this week it paid out again.
Meanwhile the parts of my work that do want a real accelerator — anything with a camera on it — are exactly the parts now sitting on a stack whose ownership changes hands in September.
The Line In That Press Release I Keep Rereading
Go back to what Microchip put in the first paragraph of what it was buying.
Not automotive design wins. Not a revenue run rate. Not a patent portfolio.
more than 100 current customers and an established developer community of more than 10,000 users
The automotive investor wrote off its money. The community got named as the asset.
Think about what that community actually is. It is people buying one HAT at a time. It is hobbyists, students, one-person integrators, somebody's weekend bird-camera project, a guy in a basement in upstate New York wondering whether his pump is dying. It is the lowest-margin, least glamorous, most support-intensive corner of the customer base. On a board deck it looks like overhead.
And when the balance sheet failed, that overhead is what made the company acquirable instead of liquidatable. Ten thousand people who already know how your toolchain works is a distribution channel that a large semiconductor company will pay real money for. Nothing else on the list held its value.
What I'm Changing
One thing, and it is small enough that I will actually do it.
I'm adding a column to the bill of materials. Not "is this vendor healthy," because I could not have answered that about Hailo and neither could you. The question I can answer is: if this vendor disappeared tomorrow, does my system keep running? Two sub-questions. Is there a drop-in second source for the part? And does anything in my build depend on a toolchain only that vendor ships?
My expectation is that every line passes, because the stack is deliberately boring — commodity sensors, current monitoring, open firmware, math I can explain out loud. But "I'm pretty sure" is not a document, and I have never actually run the check.
If it comes back clean, that table stops being an engineering artifact and becomes the most concrete promise I can make to a building owner: nothing in your system depends on a startup staying solvent. Every vendor in here has three competitors and none of them own your data or your logic.
I do not know anyone else selling small building monitoring who is offering that, and after this week I understand why it's worth offering. The people who bought a monitoring system built on somebody's clever proprietary box are, right now, reading acquisition notices and trying to figure out what happens to their support contract in September.
My customers are not reading anything. Their pumps are just running, and a very stupid smart plug is still counting the amps.
Monitoring you can explain in one sentence
Nothing gets drilled into your equipment, nothing voids a warranty, and at ninety days you get a documented baseline of what your machinery actually does — yours to keep, whether or not you stay a customer.
See how it works