The Machines Picked the Dollar

Bitcoin didn't lose the argument. It lost the first round of a fight nobody told you had started.

Every time a wallet gets drained, the same people show up to tell me Bitcoin is broken, and every time I have to explain that they are auditing the wrong ledger.

Before we start, I'm going to tell you where this ends, because two-thirds of the way down I concede a couple of things that are going to make my Bitcoin bros smack my head, and I'd rather you see them coming than think I've defected.

Here's the destination. This year three different parties went looking for money that settles without a bank — a sanctioned state, a swarm of AI agents, and the Kremlin's export sector. Not one of them chose Bitcoin. All three chose something dollar-shaped. That is a real loss and I'm not going to dress it up.

It is also the most encouraging thing that has happened to the argument in a decade, and by the end I think you'll see why.

Exhibit A ran three weeks ago and everybody read it upside down

Coinkite shipped firmware version 4.0.1 for the Coldcard in March of 2021. A build configuration error caused devices to fall back to software random number generation instead of pulling from the hardware entropy source. On the Mk2 and Mk3, effective key strength dropped from a designed 128 bits to about 40. On the Mk4, Mk5 and Q — the current hardware, on every firmware shipped before July 31st of this year — Coinkite's own estimate is about 72 bits, and independent analysis puts it as low as 32 against a targeted attacker. This ran for five years and nobody caught it. On July 30 somebody figured out how to enumerate that keyspace. Over four days and four waves they took roughly 1,816 coins, about $116 million, out of more than 5,200 wallets. The first wave was 594 coins in twenty-five minutes.

We got a solid week of drama out of it. Eulogies, I-told-you-sos, think pieces about whether self-custody was ever viable.

Now go find the part where Bitcoin failed.

A build script at one vendor, five years ago, wrote weak seeds onto a device. The protocol then did precisely what it was designed to do: it moved coins for whoever held the keys, immediately, irreversibly, at three in the morning, without asking anyone's permission and without a bank anywhere in the transaction. No compliance desk paused it. No correspondent chain slowed it. Nobody could reverse it.

That is not a bug report. That is the specification. Hold onto that sentence, because the rest of this article is about three parties who would pay almost anything for exactly that behavior.

The comparison nobody runs

$116 million is real money and I'm not making light of it. But we should probably put it in context.

Global payment card fraud runs about $33 billion a year, per the Nilson Report. That's roughly $90 million a day. Which means the entire Coldcard exploit — the four waves, the 5,200 wallets, the week of obituaries — is about thirty-one hours of ordinary, budgeted, actuarially-expected card fraud inside the traditional system.

Nobody wrote a eulogy for Visa. There was no think piece asking whether payment cards were ever viable.

And that's just the fraud. The errors are funnier. Citigroup once credited a single customer account with $81 trillion — roughly four times the annual GDP of the United States — and caught it a day and a half later. Citi also wired Revlon's lenders about $900 million of its own money by accident, and spent years in court trying to get half of it back. These are not crises. These are Tuesdays. They generate a news cycle and a consent order and everyone moves on, because the losses land in the right column of the right report and the paperwork is immaculate.

Traditional finance can lose two hundred million before lunch and it is fine, because it is documented. A hundred and sixteen million goes missing from a hardware wallet and we get a month of funeral arrangements for an asset class.

I want to be precise about my own claim, because this is where my side usually starts lying. I am not saying Bitcoin is safe. Self-custody is genuinely hard, most people are bad at it, and the Coldcard victims did everything the community told them to do and got robbed anyway.

And I'm going to make that specific, because vague contrition is just another way of not saying anything.

I wrote a book about this. In the chapter on setting up a hardware wallet, I tell you to generate a new seed phrase on the device itself, and I put a parenthesis right there on the page: we will get crazy about this later.

I never wrote the later part.

The later part is dice. Coldcard has supported it for years — you roll a physical die fifty times and the device builds your seed from your rolls. Fifty rolls supply roughly 128 bits on their own, ninety-nine gets you 256, and the whole path never touches the firmware code that broke. Anyone who rolled dice was immune to this on every model, on every firmware version, for the entire five years the bug was live.

My readers didn't get that from me. They got a parenthesis. That one's mine, and it goes in the second edition.

But notice what kind of failure that is. I gave incomplete instructions for operating a device. That is an indictment of me, and of how our whole community talks about self-custody, and we have earned it. It is not an indictment of the monetary properties, and collapsing those two things is how you end up unable to think.

What the thing was always about

Bitcoin was never about the blockchain. It was never about wallets, or being your own bank, or the number going up.

It was about one promise: no official, no banker, no committee, no memo signed at midnight can add units to the network. That's it. That's the whole product. Everything else is implementation detail.

For everyone I’ve met who truly grasped Bitcoin, the turning point was identical: the instant they realized no government can tamper with this currency. Not because they'd read a white paper. Because they'd watched 2008 and understood, in their body, that the money was a promise and the promise had a signature line on it. The General Accounting Office put the cost of that one at north of $22 trillion in lost output. Nobody's keys were compromised. Every signature on it was valid.

The idea is not remotely new. People have been proposing a currency tied to the physics of electricity for a hundred years. Edison floated it. Buckminster Fuller floated an energy denominated version. Hayek wrote The Denationalisation of Money in 1976 arguing the state should not have the monopoly at all. I wrote a book about that lineage recently and I'm not here to sell it to you (but if you're interested). The point is that the impulse predates the technology by a century. Bitcoin is not the idea. Bitcoin is the first version of the idea that shipped.

Everybody saw it. Again. Nothing happened. Again.

Three weeks ago I wrote about Washington stepping in alongside Tokyo to keep Japan from having to liquidate Treasuries — the first coordinated intervention between the two since 2011, staged specifically so nobody would find out what the bond market would do. It got a day of coverage and then everyone moved on, because there was no crash to photograph.

On August 19th and 20th, the Treasury announced it is doubling its long-end liquidity support buybacks. Maximum size per operation goes from $2 billion to at least $4 billion, covering the 10-to-20 year and 20-to-30 year sectors, effective September 9th and running through November 4th.

The government is buying its own debt to hold the yield down on the debt it is still issuing.

Now read Treasury's stated reason, which is my favorite sentence of the month: they cite "consistent strong sponsorship from market participants" and "the significant volume of high-quality offers Treasury routinely receives in longer-dated buyback operations."

They doubled the intervention because demand is excellent. Everything is fine. We are buying more of our own bonds because people love our bonds so much.

That same week the national debt crossed $40 trillion.

And here's the part my side will skip, so I'll do it myself: it didn't work. Yields fell on the announcement and then rebounded the next day, wiping out the decline. Analysts put out notes about limited relief. This was not a rescue. It was a sedative, administered to a patient who is going to wake up.

The price did what the price does

Bitcoin went from about $71,000 to $77,300 as of this morning, opened 5.4% above yesterday, and depending on which exchange and which minute you screenshot, somebody is writing a headline that says $79,000. Ten grand in a matter of days.

I have watched the bitcoin tag all week and it is nothing but that number. "Bitcoin Just Added $10K in One Day." "Bitcoin Just Broke $79,000." There is a publication that as of an hour ago was still telling readers we're holding near $64,000, which is only wrong by twenty percent and four days.

The price is the least interesting thing on this page. A liquidity injection moved a liquidity-sensitive asset. That is not a monetary transition, it is a reflex, and if you are taking the reflex as vindication you're going to take the next drawdown as refutation, and you'll have learned nothing in either direction.

Two states already ran the experiment

While we were arguing about firmware, two governments under real pressure went and tested the thesis, and neither result is the one my side puts on the slide.

On July 22nd, Russia passed a law legalizing crypto trading. The headline went around fast. Read the statute and it deflates. Cryptocurrency remains prohibited as a means of payment inside Russia. Retail investors have to pass a knowledge test. Non-qualified investors — about 98% of the country — are capped at 300,000 rubles a year, roughly $3,840. Banks are required to reject transfers to unlicensed service providers. Individuals cannot move funds to their own wallets. Providers have until July 1, 2027 to get licensed by the Central Bank, and only assets above a 5 trillion ruble market cap qualify automatically, which as of today means Bitcoin, Ether and Tether.

That is not adoption. That is a cage with a Bitcoin-shaped door, and it is worth understanding precisely, because half the people who will share this article are going to call it a win.

But read the exemption, because the exemption is the confession. Foreign-trade companies can use crypto for cross-border payments without restriction. So: the citizen gets a $3,840 annual allowance, a quiz, and no keys. The exporter gets unlimited sanctions-resistant settlement. Russia did not legalize Bitcoin for its people. It legalized it for its balance of payments, and it built the cage to make sure the two never touch.

Then there's Iran, which is the more interesting case by a mile.

Since April, Tehran has been charging a toll for transit through the Strait of Hormuz — a dollar per barrel, up to about two million dollars a ship, denominated in crypto. The mechanics, first reported by the FT, are almost comically modern: the vessel emails its cargo specs, Tehran verifies the documents and issues an invoice, and the captain has seconds to complete the transfer before the sanctions exposure catches up with him. Estimates run to $20–21 million a day.

Here is the part my side leaves out. Both Chainalysis and TRM Labs looked at it and concluded the settlement is happening in stablecoins, not Bitcoin, whatever the invoice says.

I'm not going to pretend that's a victory. It isn't. But sit with what Iran actually is for a second, because it's about to matter.

Iran is an entity that cannot use correspondent banking. It needs finality in seconds. It operates at three in the morning across jurisdictions that will not clear its payments, with counterparties it has no legal relationship with and no ability to sue. It has to settle small, fast, and irreversibly, over and over, with no institution standing behind any of it.

Hold that list.

Here's the part I actually wanted to tell you

AI agents are starting to move money. Not "will." Are. And they do not have bank accounts.

Go back and reread the Iran list. Cannot use correspondent banking. Needs finality in seconds. Operates at three in the morning across jurisdictions that won't clear it, with counterparties it can't sue. That is not a description of a rogue state. That is a description of a piece of software.

An agent cannot pass a KYC interview. It cannot hold a Fed master account. It has no legal personality, no branch, no signature, no jurisdiction. When it needs to buy an API call, rent compute, pay another agent for a subtask, or settle forty thousand transactions of eleven cents each in an afternoon, the entire twentieth-century apparatus of correspondent banking, batch settlement, business hours and chargeback windows is not a slow option. It's not an option.

What it needs is money that is programmable, final, permissionless, available at three in the morning, and — this is the constraint everybody underrates — stable enough not to drift while the work is in flight. An agent quoting a price, doing the job, and settling ninety seconds later cannot be exposed to a four percent move in the middle.

So here is the honest scoreboard, and my side is not going to like it.

The machines picked the dollar. The agentic payment rails being built right now — x402, AP2, the whole emerging agentic settlement stack — are overwhelmingly stablecoin rails, and stablecoins are dollar-denominated. Given a free choice of monetary substrate, unburdened by ideology, with no laser eyes and no priors, the machines chose tokenized dollars. Bitcoin's volatility disqualified it for exactly the reason above. That's a loss. Write it down as a loss.

And notice that Iran, facing the identical constraint set with vastly more motivation to embarrass the dollar, made the identical choice. When the pressure is real and the deadline is measured in seconds, everybody picks the stable unit. Everybody. Ideology is a luxury good and it does not survive contact with a settlement window.

But look at what the machines actually rejected, because it is not what you think. They did not reject hard money in favor of the Federal Reserve. They rejected banks. They routed around the entire permissioned, hours-bound, identity-gated settlement system that Nixon's dollar has lived inside since 1971, and they did it in about eighteen months, for engineering reasons, with no manifesto.

The stablecoin is a compromise object. It gives you dollar stability on crypto rails, and in exchange it hands one private issuer the power to freeze your balance, blacklist your address, and honor a subpoena on Tuesday. Which is to say: for the first time, the machines have made "who can switch off your money" into a live engineering question rather than a political one. It's now a line item in an architecture doc.

And once that question is on the table, it does not stay answered. Because the thing under the stable unit still has to be something. Today it's Treasuries — the same instrument the government is now doubling its buybacks of, sitting on the same $40 trillion tab, in the same jurisdiction where an attorney general can decide by memo who does and doesn't get audited. The machines built a beautiful frictionless settlement layer and collateralized it with the exact liability I spent three thousand words on last month.

That's not a gotcha. That's just the position. Volatility is a solvable problem — it gets solved with duration, with hedging, with instruments that don't exist yet. Issuer discretion and sovereign solvency are not solvable problems. They're structural. Bitcoin lost this round on the axis that engineering fixes and won on the axis that engineering can't.

So

Currency regimes turn over. It's happened repeatedly across twenty-five hundred years, and the last time was within living memory — Bretton Woods, 1944, the dollar taking the crown off the pound while Britain was still technically winning a war.

Is Bitcoin the final move? I don't know. Anybody who tells you they know is full of shit, and I include the version of me that gets excited on a green week.

What I'm confident of is narrower. Three constituencies showed up this year wanting the same shape of thing — a sanctioned state, a sanctions-exposed exporter, and a swarm of software — and not one of them wanted Bitcoin, specifically. All three wanted settlement that doesn't route through a bank. The demand is no longer only human and no longer only ideological. There is now a second constituency for money that settles without permission, and it is made of software, it doesn't read Hayek, it has no politics, and it is growing faster than the human side ever did. It picked a dollar-shaped answer first. It picked a bank-free one immediately.

I'm not asking you to believe a word of this. I'm asking you to put down whatever you were certain about before you started reading and go look at the agentic payment stack yourself, in the quiet of your own solitude, and ask what it's going to want in five years when someone finally builds a unit that's both stable and unownable.

If you want the human version of the argument, read Lyn Alden, or Jeff Booth, or Jack Mallers. If you want the intellectual version, read Hayek. If you want the machine version, nobody's written it yet.

Yes, there's an AI watermark on this one too. I'm dyslexic; the machine does my orthography the way glasses do somebody's astigmatism. If you're focused on wordcraft, your missing the point.

Sources: Coldcard exploit — TRM Labs, CoinDesk, The Hacker News; per-model entropy figures and the dice-roll remediation from Casa's vulnerability analysis and Coinkite's own documentation. Card fraud — Nilson Report. Citigroup errors — CNBC, Bloomberg. 2008 crisis cost — GAO-13-180 (2013), via Better Markets. Treasury buybacks — U.S. Treasury press release SB0607, Aug 19–20 2026; CNBC on the yield rebound. Russia — The Moscow Times, July 22 2026. Iran — Financial Times via Decrypt and The Block; Chainalysis and TRM Labs on stablecoin settlement. Prices as of 9:08 a.m. ET, August 21, 2026.