The number in that title is real in a single sense: it circulated. We are going to tell you where it comes from before we use it, because where it comes from is the story.
On June 3, 2026, Amazon switched off Amazon One, the palm-recognition system that let you pay for groceries by hovering your hand. The trade press recorded a farewell statistic: over the prior year, the system had processed 36.7 million items across 17.7 million shopping sessions. That pair of figures now serves as the product's public epitaph. It appears in Mass Market Retailers on January 28 and in ID Tech Wire on January 29, in near-identical sentences, unattributed in both. Our research made three separate passes at finding the original source. We found outlets repeating the sentence. We never found who first said it. Amazon's own FAQ page for the product resolves to the AWS homepage.
Here is the one usage figure Amazon has actually put its name to: Amazon One had been used more than 8 million times as of March 2024, a company statement carried by TechCrunch and Amazon's newsroom. That is a lifetime total, at a different order of magnitude from the anonymous annual figure in the obituaries.
This essay is about the unit economics of a dead biometric, and the first fact of those economics is that there are no public units: no hardware cost per reader, no enrollment cost, no per-transaction economics, ever. A payment product lived nearly six years inside one of the largest retailers on earth and died without one audited number escaping. What follows is what can be reconstructed from the numbers that did escape, starting with a step nobody who repeated the epitaph appears to have taken: dividing one of its numbers by the other.
On January 27, 2026, Amazon announced it was closing its physical Amazon Go and Amazon Fresh stores. All 72 US locations went: 57 Fresh supermarkets and the 15 remaining Go convenience stores, most shut by February 1, California stragglers following weeks later under state labor-notification rules. Amazon's stated reason was about the stores, not the biometric: the company said it had not "created a truly distinctive customer experience with the right economic model needed for large-scale expansion."
Four months later, on June 3, Amazon One went dark at retail: readers removed, service ended, and, per the trade reporting, enrolled palm data deleted. The stated reason: limited customer adoption. Healthcare check-in, where the palm verifies patients at institutions like NYU Langone Health, was exempted "until further notice." GeekWire's framing tied the two events together in its headline: Amazon was ending its palm system "as it closes physical stores."
The host format died first; the gesture followed four months later. That timeline tempts a simple story: the palm did not fail, it was orphaned. Readers in lanes Amazon owned, in a format Amazon abandoned; when the stores went, the biometric had nowhere to stand. There is truth in that story, and one fact ruins its simplicity.
Amazon One was not confined to the 72 stores that closed. Amazon announced it was rolling the palm readers out to all 500-plus Whole Foods Market stores in the United States, and the announcement is still on its newsroom site. That is, as far as we can establish, the widest deployment any payment biometric has ever had in American retail: every lane of a national grocery chain, installed by an owner who controlled both the hardware budget and the checkout software, with a loyalty program already wired to the same account.
And Whole Foods is not a dying host: in the same announcement that closed Fresh and Go, Amazon said the chain's sales are up more than 40 percent since 2017, that it has more than 550 locations, and that it plans more than 100 new ones.
Now hold those two facts against the standard theory of why payment networks fail, the two-sided threshold: merchants will not install readers until enough users enroll, users will not enroll until readers are everywhere, and the product starves in the gap. Amazon One may be the cleanest falsification of that theory we can point to, because Amazon bought out one entire side of the market. The merchant was the parent company. Distribution reached functionally 100 percent of a thriving 500-store chain, the condition every payments startup claims would unlock the flywheel.
The flywheel did not turn. The product died citing limited customer adoption anyway. Density was achieved on the side money could reach; the side made of people deciding whether to register a palm print with Amazon to save a few seconds never showed up. Whatever killed Amazon One, it was not a chicken-and-egg problem. The eggs were provided at corporate expense.
Take the obituary statistic at face value and do the arithmetic: 36.7 million items across 17.7 million sessions is 2.07 items per session.
That is not what grocery shopping looks like. A payment method earning its keep at the checkout lanes of 500 supermarkets should show baskets in the tens of items; two items per session is a convenience-store grab or a stadium concession run, a beer and a pretzel. Three readings of the number are possible, and every one is bad news for the palm.
First reading: the figure does not describe Amazon One at all. "Items processed" is a strange metric for a payment gesture; it is the native metric of Just Walk Out, the camera-and-sensor system that counts what you carry. A cashierless-store statistic may simply have been attached to the palm reader somewhere in the retelling and repeated outlet to outlet, referent unchecked. The verbatim, unattributed repetition is exactly what that would look like.
Second reading: the sessions are real palm sessions, but they lived overwhelmingly in the small-basket formats, the Go stores where a palm opened the gate, the stadium and airport shops. The arithmetic is coherent: seventeen million annual sessions concentrated in a few dozen high-traffic sites is hundreds of scans per site per day, genuinely healthy usage. But those are precisely the venues that closed. On this reading, the palm's demonstrated demand lived almost entirely inside the host that died, and the 500 thriving grocery stores contributed a rounding error.
Third reading: the sessions really do include meaningful Whole Foods checkout use, in which case the items-per-session figure is simply wrong somewhere, because the division does not describe grocery baskets.
We cannot determine which reading is true, and neither, apparently, can anyone else, because nobody who published the number said what it counts. That is the point. The one statistic history kept for this product either describes a different product, describes usage concentrated in the venues that were shutting anyway, or fails its own arithmetic. An epitaph that dissolves under division is not evidence the product was working. It is evidence nobody was ever shown the real ledger.
Run the same exercise on the figure Amazon did put its name to: more than 8 million uses, lifetime, as of March 2024. The product launched in late September 2020. Be maximally generous to the palm and pretend all 8 million uses happened in a single year entirely at the 500 Whole Foods stores, ignoring every other venue. That still averages roughly 44 scans per store per day. A single checkout lane at a busy supermarket serves more customers than that before lunch. Distribute the uses across the full three-and-a-half-year window and the whole national footprint, and the number falls to something like a dozen daily uses per venue.
Every assumption there is stated, every one bends in the product's favor, and the result is still a per-store volume a spare cashier could absorb without noticing. This is what limited customer adoption means when you convert it from a corporate adjective into a rate.
None of this required inside information. In October 2023, the analyst firm VDC Research looked at the palm reader and wrote that it "looks like revolutionary payment technology but it's only a small improvement in retail checkout," noting that initial usage at Whole Foods was minimal and that the enrollment step, handing a biometric to Amazon for a marginal convenience, was the friction that would decide the outcome. That note, the closest thing to a public unit-economics analysis of Amazon One, contains no unit costs either; none were available. Its structural call aged perfectly, and the wall was visible from outside two and a half years before the shutdown: the product asked users for something priceless and offered seconds in return.
The clean way to see the episode is to split what Amazon built into three layers and follow each to its fate.
The store format, Go and Fresh, is dead: all 72 US locations closed in February 2026, on Amazon's own economics, in Amazon's own words.
The identity gesture, the palm, is dead at retail: switched off June 3, readers removed, data deleted, four months after its flagship venues closed and despite its 500-store second home surviving in excellent health.
The mechanism, Just Walk Out, lives: Amazon now offers it to third parties, and its own announcement of the store closures pointed to concession stands at sports stadiums as the future. We covered that afterlife in a companion essay, Amazon Is Closing Every Store It Built Just Walk Out For. The Same Week, It Called It the Future. The short version is that the technology outlived the store it was invented to prove, and is sold now to operators who bring their own venues.
A fourth data point hides in the exemption: the one place the palm itself survives, healthcare check-in, is the one place it is sold as an institutional service to a buyer with its own premises and its own reason to pay. NYU Langone adopted it through AWS for patient check-in.
Put the survivors side by side and the pattern is not subtle. What lived found a buyer outside Amazon, with an installed base of its own, in a use case where the customer does not have to be recruited one palm at a time. What died needed Amazon to keep funding the venue, or needed millions of individuals to enroll a body part for a convenience they had been declining for years. The demo layer died first and most completely. The layers with paying institutional customers are still here.
That is the corrected version of this essay's premise. The gesture that sells a demo and the layer that clears a market are different products, but what separates them is not network density; Amazon proved you can buy all the density one side can hold. It is that a demo layer never acquires customers who would fight for it, and when the strategy carrying it changes, no constituency is left to argue. The stores had a P&L argument and lost it. The palm never got a hearing: nearly six years in, there was nothing on its side of the ledger anyone outside the company had seen.
The three things worth carrying out of this all reduce to one: somebody should have been writing the numbers down while the product was alive.
Start with the side of the market you control, because controlling it tells you nothing about the side you do not. Amazon owned the lanes and drove merchant adoption to 100 percent by fiat. The metric that mattered was scans per store per day, and it was knowable from year one. If your product's success depends on end users volunteering something, instrument that voluntary rate first and believe it early. Everything you can mandate is a vanity metric precisely because you can mandate it.
Then ask which layer has a buyer, and specifically whether that buyer is an institution or a recruitment campaign. Both survivors here, Just Walk Out at stadium concessions and the palm at hospital check-in desks, are sold to institutions that bring their own venues and users. Every casualty needed either continued patronage from the parent's strategy or one-at-a-time consumer enrollment. If your roadmap's most magical layer is also its most recruitment-dependent layer, plan for the magic to die first, and put a number on what would justify keeping it alive.
And publish your units while you are alive, internally at least and externally if you can. Amazon One died without ever stating a hardware cost, an enrollment count, or a per-store usage rate, and so its public record is now an unattributed sentence whose arithmetic describes a two-item basket. A product that never shows its ledger does not avoid being judged. It gets judged anyway, posthumously, by whatever number happens to be circulating. If the only statistic that survives you misdescribes you, that is not the trade press's failure. It is what silence buys.
Seventeen point seven million sessions, whatever they were, was not enough. Not enough provenance to trace, not enough grocery in the basket math, and not enough of a constituency to save the product it eulogizes. The palm reader was a handshake offered to the market for nearly six years. The market, politely and at the widest distribution a retail biometric has ever enjoyed, kept its hands in its pockets.
What survives you is whatever number is circulating
Amazon One ran nearly six years and died without one audited number escaping, so its public record is an unattributed sentence whose arithmetic describes a two-item basket. A product that never shows its ledger does not avoid being judged; it gets judged posthumously, by whatever figure happens to be circulating. The same hole opens under any system whose work is summarised after the fact rather than recorded as it happens. Chain of Consciousness is the ledger written while the work runs, a tamper-evident record of what an agent did, on what inputs, in what order. It will not make an agent right. It means the numbers describing it come from the run.
Hosted Chain of Consciousness · Verify a record
pip install chain-of-consciousness · npm install chain-of-consciousness