The Pin

---

The Humane AI Pin was supposed to be the iPhone of artificial intelligence — the device that ushered AI off the
screen and into the world, projected onto your palm, summoned by your voice, freeing humanity from the
addictive glass rectangle. It was founded by ex-Apple luminaries, lavished with venture capital, and unveiled
with the kind of reverent staging usually reserved for products people actually end up buying. It raised about
$230 million. And then it met the market.

The reviews were not mixed. They were a massacre. The Pin was slow. It was unreliable. It overheated. Its
battery died. Its party trick — projecting a laser display onto your hand — was a gimmick that worked badly in
any real lighting. It cost $699 plus a $24-a-month subscription to be worse than the phone already in your
pocket at every single task it attempted. By 2024 it was being returned faster than it was sold — a detail
worth pausing on, because returns exceeding sales is not a soft launch or a slow start. It is a market actively
rejecting a product, physically mailing it back. Humane shipped, by most accounts, fewer than ten thousand
units. In February 2025 it sold itself to HP for about $116 million — roughly half what it had raised — and
HP discontinued the Pin essentially on contact. The company's name, fittingly, is now followed in databases by
the word "deceased."

This essay is about why that happened, why it happened again to the Rabbit R1 in almost exactly the same
shape, and why — with both corpses still warm — OpenAI and Jony Ive are spending billions to do it a third
time. It is about the single iron law these companies keep breaking, and about what it tells you when the
smartest, best-funded people in technology keep running at the same wall because the wall is shaped like an
enormous market they are sure must exist.

The law of the rectangle

Start with the law, because every device in this graveyard died of the same cause and the cause is not
mysterious. It is the most successful consumer product in human history.

The smartphone is not a gadget. It is a near-perfect convergence device — a camera, a map, a wallet, a
messenger, a browser, a music player, a flashlight, and, increasingly, a fully capable AI assistant, all in one
slab that a few billion people already own, already carry, already understand, and have already paid for. Any
new hardware category that wants to live in a person's pocket or on their lapel must answer one brutal question:
what do you do that the rectangle does not already do better? And the consistent, fatal answer from every AI
gadget so far has been: nothing. They do a subset of what the phone does, worse, on a device you have to buy
separately, charge separately, carry separately, and learn separately.

This is why the Humane Pin failed and it is why it was always going to fail. Its core proposition was that you
would abandon the most refined consumer product ever made in favor of a slower, dumber, single-purpose puck.
That is not a hard sell; it is an impossible one, because it asks the customer to trade down. The AI inside the
Pin could, in principle, run on the phone the customer already owned — and now, increasingly, it does, because
Apple and Google have simply absorbed the assistant into the device everyone already carries. The moment the
incumbent platform adds your one feature, your entire reason to exist as separate hardware evaporates. The pin
was a feature wearing the costume of a product.

The same movie, in orange

If Humane were the only case, you could call it a fluke of bad engineering. It is not the only case. The Rabbit
R1 ran the identical arc, just in a cheerier color, and the parallel is the proof that the problem is structural
rather than incidental.

The R1 was a small orange box — designed, like the Pin, by a celebrated design house — that promised a
"large action model" capable of operating your apps for you. It was cheaper than the Pin and it sold better at
first: about 100,000 units in pre-orders, a genuine flash of consumer enthusiasm. And then the same market
verdict arrived, just slightly delayed. Five months after launch, the R1 reportedly had about 5,000 active
users left — a 95% abandonment rate. Ninety-five percent of the people who were excited enough to pre-order
the thing had put it in a drawer. The company was later reported to be struggling to make payroll.

And then came the detail that turned the R1 from a failure into a punchline. A developer discovered that the
device's operating system was essentially a modified version of Android — that the "revolutionary" AI
gadget was, underneath, an Android app on cheap hardware, and that you could install the very same "Rabbit
Launcher" experience on a phone you already owned. The entire physical product was, functionally, an app that
had been given a plastic body and a $199 price tag. It was the law of the rectangle made literal: the thing the
gadget did could be done by the rectangle, because the gadget was the rectangle, wearing a costume, sold
separately.

Two companies. Two celebrated design pedigrees. Two enormous waves of hype. Two near-total market rejections,
for the identical underlying reason. At this point the pattern is not an accident. It is the category.

The six-and-a-half-billion-dollar encore

Which brings us to the part that should make a forensic reader sit up, because the appropriate response to a
category that has produced only graveyards is caution, and the actual response from the most powerful company in
artificial intelligence has been to bet more on it than anyone ever has.

In May 2025, OpenAI acquired io Products — the hardware startup co-founded by Jony Ive, the designer of the
iPhone itself — in an all-stock deal valued at roughly $6.5 billion. It has confirmed plans to debut its
first consumer device in the second half of 2026: a screenless, AI-native object, always listening, designed to
be a "peaceful and calm" alternative to the addictive phone. Multiple form factors are reportedly in
development — an earbud, a pen-like device — and, most strikingly, the contract manufacturer Foxconn is said to
be tooling up to build forty to fifty million units.

Sit with that production figure, because it is the entire thesis and the entire risk in one number. Forty to
fifty million units is not a cautious pilot. It is iPhone-scale ambition for a product category whose entire
commercial history to date consists of fewer than ten thousand Humane Pins and five thousand surviving Rabbit
users. OpenAI is not dipping a toe; it is committing to manufacture, in its first generation, several thousand
times the total installed base that the category has ever successfully sustained. The bet is that the previous
failures were failures of execution — bad battery, bad chip, bad software, second-rate design — and that with
the iPhone's designer, OpenAI's models, and Apple-scale manufacturing, the category finally works.

Maybe. But notice that this is precisely the argument every doomed challenger in every graveyard has always
made: the others failed because they did it badly; we will do it well. It is the demonstration-versus-
deployment fallacy in its purest hardware form. The thing that killed Humane and Rabbit was not primarily that
their execution was poor — though it was. It was that the category itself asks the customer to carry a second
device to do a subset of what the first device already does. Better execution does not dissolve that problem. A
beautifully made, perfectly engineered, Jony-Ive-designed device that does a fraction of what the phone in your
other pocket does is still a device that does a fraction of what the phone in your other pocket does. Polish is
not a category.

The assistant already moved in

There is a development that makes the hardware bet harder than it was even when Humane launched, and it is the
thing the gadget-makers least want to discuss: the AI assistant they are selling as the reason to buy a new
device has, in the interim, simply moved into the phone.

When Humane conceived the Pin, there was a plausible-sounding gap: the phone was a dumb slab of apps, and a
voice-first AI companion was a genuinely new thing you could not get there. That gap has closed. Apple and
Google have spent the last two years stuffing increasingly capable assistants directly into the devices
everyone already carries — Gemini on every Pixel and Android phone, a reworked Siri and on-device models on
every iPhone, AI woven into the camera, the keyboard, the search bar, the operating system itself. The single
feature that was supposed to justify a separate $699 puck is now a free, pre-installed, default-on capability of
the hardware in your pocket. This is the cruelest dynamic in consumer technology and it is utterly predictable:
the platform owner watches a startup prove that a feature is desirable, and then the platform owner adds the
feature, at zero marginal cost to a device the customer already owns, and the startup's entire reason to exist
is absorbed. Humane spent $230 million to run that experiment on Apple and Google's behalf. The Pin's most
durable legacy may be that it helped convince the incumbents the assistant was worth building in.

So the AI-hardware company of 2026 is not competing against the phone of 2023. It is competing against a phone
that has already eaten its core feature and is giving it away for free. That is a materially worse starting
position than the one that produced two graveyards, and it is the position OpenAI's device will be born into.

Why the money keeps coming

If the category has a 0% success rate, the obvious question is why serious capital keeps pouring into it — and
the answer reveals something about how venture economics manufacture these doomed bets on a schedule.

The prize, if a new personal-computing category ever does open, is so enormous that it warps everyone's
judgment. The smartphone created the most valuable companies in history; "the next iPhone" is the single most
seductive phrase in technology investing, because the one time the bet paid off it paid off more than any bet
in the history of commerce. So the math a venture investor runs is not "what is the probability this device
succeeds" — it is "the probability is low, but the payoff if it works is a trillion-dollar platform, so the
expected value justifies the check." This is a defensible portfolio logic and it is also exactly the engine that
guarantees a steady supply of beautifully funded AI-hardware corpses. Each one is a lottery ticket bought
against the memory of the one ticket that hit. The funerals are a feature of the model, not a bug in it.

And there is a softer incentive layered on top: the device is a story, and stories raise money. An AI lab that
ships only a chatbot is a software company with a usage-based revenue line and an enormous compute bill. An AI
lab that is building a revolutionary physical device designed by Jony Ive is a company inventing the future of
human-computer interaction — a narrative that supports a far larger valuation, a far grander fundraising round,
and a far more thrilling recruiting pitch. The hardware does not have to succeed to do useful work for the
company that announces it; it only has to be plausible long enough to anchor the next round. That is not an
accusation of bad faith. It is an observation that the incentives reward the announcement nearly as much as the
product, which is precisely the condition under which difficult products get built on hope.

The wider overreach

The pin and the box are the cleanest cases, but they are part of a broader pattern of consumer-AI products
launched into demand that was assumed rather than demonstrated — and the assumption keeps proving expensive. The
same period saw AI features bolted onto everything, AI subscriptions pushed at customers who had not asked for
them, and flagship consumer-AI launches that reportedly bled cash at extraordinary rates as usage outran any
plausible business model. The connective tissue is identical to the hardware story: a technology genuinely
impressive in a demonstration, projected by its builders into a mass-market product the public had not actually
requested, funded on the conviction that the demo's magic would automatically convert into durable consumer
behavior. Sometimes it does. Far more often, the magic is real and the behavior never forms, because the
customer tries the thing, finds it does not improve a life already served by tools they own, and quietly stops.
The graveyard of AI hardware is only the most physical, most photographable wing of a larger building.

What would have to be true for OpenAI to win

Be fair, because OpenAI is not Humane, and the differences are real and worth stating honestly.

For the encore to succeed, a few things have to be true that were not true before. The device cannot try to
replace the phone — and reporting suggests OpenAI knows this, framing it as a new category rather than a
smartphone-killer, which is the single most important lesson it could have drawn from the graveyard. It would
have to find a genuine job that the phone does badly — always-on ambient capture, hands-free continuous
assistance, a fundamentally different interaction model — rather than a job the phone does well. It would need
the AI itself to have advanced far enough that a voice-first, screenless interface is actually pleasant rather
than frustrating, which it has not been on any prior device. And OpenAI has two things its predecessors lacked:
frontier models it controls end-to-end, and the distribution and brand to make a new behavior aspirational
rather than ridiculous. If anyone can will a new category into existence, it is the company with the most
capable AI and the designer of the most successful device of the century.

But "if anyone can, they can" is not the same sentence as "it will work," and the base rate is brutal. The
honest framing is this: OpenAI is making the most credible version of a bet that has, so far, a 0% success rate
and a 100% capital-incineration rate, at the largest scale anyone has ever attempted it. The credibility is
real. The base rate is also real. And forty to fifty million units is a great deal of aluminum to manufacture
on the strength of "this time the designer is better."

The kicker

There is a tell that runs through the entire AI-hardware graveyard, and it is the tell of every product built by
people who fell in love with the demo. Each of these devices was conceived as the answer to a question the
customer was not asking. Nobody clipped on a Humane Pin because they hated their phone; they bought it because a
beautiful keynote told them the future looked like this, and then they used it for a week and remembered that
the future was already in their pocket and it had a screen and it worked. The graveyard is full of solutions in
search of the problem, and the problem they keep failing to find is that the problem was solved, completely, by
a device a few billion people already own. The hard truth the category cannot metabolize is that "the AI should
live somewhere other than the phone" is a designer's wish, not a customer's need — and customers, in the end,
buy needs. They have demonstrated this twice now, decisively, with their wallets and their return-shipping
labels, and the industry's response has been to assume they will demonstrate it a third time in the opposite
direction if only the aluminum is prettier. They will not. The verdict is in; only the defendants refuse to hear
it.

Humane raised two hundred thirty million and shipped ten thousand pins into a wall. Rabbit sold a hundred
thousand boxes and kept five thousand users. OpenAI looked at both graves, picked up a shovel, and ordered fifty
million units. Somewhere a designer sketches a screenless object of great beauty, and in the customer's other
pocket, glowing faintly, the rectangle waits — patient, capacious, and already paid for.