Fifty Thousand Copies of One ID

Bitcoin ended the week 54 percent below its record. In the same 48 hours an AI scanner found a hole in its networking layer and the fix reached the relays in a point release. Those are two different clocks, and only one of them says anything about the next ten years.
Fifty Thousand Copies of One ID

Fifty Thousand Copies of One ID

A vault door covered in a thousand identical keyholes, one of them lit

The loud number and the quiet one

On Saturday morning bitcoin printed $57,549. Earlier it had touched $57,655, under its summer low and, as CoinDesk put it, below $57,000 for the first time in two years. That number is loud. It is on every feed, it will be screenshotted into a thousand posts, and it is why your phone was busy before breakfast.

The quiet number is 50,000. An inv message on the bitcoin peer-to-peer network can carry up to fifty thousand entries, and until August the software did not skip duplicates. An attacker could announce one orphan transaction with thousands of missing parents, then repeat its identifier fifty thousand times in a single message. Each such message stalled the receiving node “for over a dozen seconds, during which it processes no messages from other peers, including blocks.” The advisory went public on Friday as CVE-2026-117348, rated medium severity and, unusually, “easily discoverable.”

Both things landed in the same news cycle. Only one of them tells you anything about the next ten years.

Here is the claim I want to defend. What fell this week was not bitcoin’s money. What fell was a rental agreement between bitcoin’s price and the credit market, and the rent just got repriced. The protocol underneath kept producing blocks, and one of its doors quietly got a new lock. If you hold bitcoin for the second thing, this week cost you nothing you did not already agree to pay in volatility. If you hold it for the first, you own a levered position on the Federal Reserve and on the AI capital cycle, and this week you found out how levered.

What actually sold

Start with the composition of the selling, because it is not mysterious and none of it is about bitcoin’s properties.

The first flow was leverage. About $1.53 billion of crypto positions were liquidated in 24 hours, $1.4 billion of them longs. Bitcoin accounted for roughly $540 million of that, ether for a good part of the rest as it fell more than 8 percent toward $1,600.

The second flow was funds. I went through the daily US spot bitcoin ETF table myself for the fifteen trading sessions from 2 to 20 November and added up the totals. Net redemptions came to roughly $3.59 billion. One fund, BlackRock’s IBIT, accounted for $2.94 billion of it, about 82 percent of the month. On 19 November the complex lost $582.3 million in a single day, $541.1 million of that out of IBIT, its largest one-day outflow on record. The next day another $124.5 million left.

Daily net flows into US spot bitcoin ETFs, 2 to 20 November 2026

The third flow was not a flow at all. It was a repricing of everything that had been bought as a growth asset. The Nasdaq fell 5.28 percent, the S&P 500 3.70 percent, the Nikkei 5.06 percent. Nvidia lost about 12.5 percent. The miners, who had spent two years telling investors they were really AI data center companies, discovered what that sentence costs when the AI trade wobbles: MARA down 17 percent, IREN down 17 percent, CleanSpark down 16 percent. They were levered twice, once to the hashprice and once to the narrative they had borrowed. Those index percentages are my own arithmetic on 20 November closing prices.

Now look at what rose. The dollar index closed near 103.84, up on the day. And look at what fell that was supposed to protect you: gold down 4.62 percent, silver down 7.41 percent, copper down 2.64 percent.

That last set is the part worth sitting with. In a week that was ostensibly about the debasement of paper money, both of the assets people buy to hedge debasement fell hard, and the currency they were hedging against went up.

Silver, the Nasdaq, the Nikkei, gold, the S&P 500, copper and the Dow on 20 November 2026

The arithmetic under the AI trade

To understand why bitcoin was in that basket, you have to look at what basket it was.

On Thursday, Anthropic completed the largest initial public offering in history: $81.3 billion raised, beating SpaceX’s $75 billion from June. The stock was priced at $83, opened at $111.50, and closed its first day at $97.17. On Friday it fell about 12 percent to roughly $86, a stone’s throw above the offer price. A company valued at roughly $1.5 trillion gave back about 12 percent of that in its second session.

The interesting document is not the prospectus. It is the arithmetic that FT Unhedged ran, and which Latent Space reported on Thursday. Take a 20 percent return on roughly $1 trillion of compute commitments. Apply a 40 percent operating margin and a 17 percent tax rate. The revenue required is about $602 billion a year. That is more than Alphabet’s $471 billion. More than Apple’s $478 billion. It is two-thirds of the $908 billion earned last year by all seventy-six software and IT companies in the S&P 1500 combined.

That is not a bear case against artificial intelligence. It is a statement about what has to be true, within a few years, for the current price of the compute buildout to be justified. When a number like that has to arrive on schedule, the discount rate stops being a technicality and becomes the whole story. And the discount rate is not set by engineers.

A quarter-point increase in the federal funds rate in December is currently priced at more than 98 percent. Not a cut. An increase. Treasury Secretary Scott Bessent spent Friday pushing back on calls for a half-point move, which tells you the debate is between a small tightening and a large one. On 19 November the ten-year closed at 5.04 percent against a two-year at 4.76 percent, a curve of about 28 basis points. On Friday the long end fell another 10 basis points to 4.95 as money ran for cover, while the front end stays pinned by the coming hike. That is a curve flattening from below, and it is not the shape of an economy about to accelerate.

This is where the Austrian reading earns its keep. Credit priced below the market’s time preference does not create nothing. It creates real things in the wrong order: too many data centers before the demand that would pay for them, too many corporate treasuries stuffed with bitcoin before the buyers who would hold it, too many miners financing ASICs against a hashprice that assumed the last cycle’s margins. The liquidation cascade is not the cause of the damage. It is the accounting by which the damage gets recognized. Roughly $1.5 billion of positions died in a day because they had been built on the assumption that the rate would keep falling. It did not.

Bitcoin got swept up because it had been sold to the same marginal buyer as the AI trade, using the same pitch. Not “here is a fixed schedule of issuance that no committee can revise,” but “here is the asset that goes up when money gets easier.” Those are two different products. In 2020 and again in 2024 they traded as one. This month they were separated at the till.

The other clock

While the price was doing that, the block interval was ten minutes.

That sentence is easy to write and easy to skip past, so let me make it concrete. No committee met to decide whether blocks would continue. No treasury official was consulted. No exchange could halt the chain to protect a price. The difficulty adjusted on schedule. Nobody’s balance changed because somebody else’s did. And a few thousand people who run nodes checked, without being asked, that the rules they had been enforcing for years still held.

Then the second thing happened.

Bitcoin Core keeps transactions whose parents it has not yet seen in an “orphanage,” and asks peers that announce an orphan for the missing parents. Since version 29.0, when a peer announces an orphan by identifier, the node builds the list of that orphan’s parents and checks each one against its mempool and against filters of recently seen transactions, to decide whether to ask the peer for them. It is sensible design: the node is reconstructing a chain it does not have. What it did not do was skip duplicate entries inside a single announcement.

So: send an orphan with thousands of missing parents. Then announce its identifier fifty thousand times in one message. The node performs the same expensive check for every copy, and for a dozen-plus seconds it is deaf — to transactions, to blocks, to everything. Send more messages and you can keep it deaf.

The economic consequences are not abstract. An exchange running an affected node could be slow to credit deposits, which is bad for the exchange and worse for the customer staring at a confirmed transaction and an empty balance. A miner running an affected node could spend real electricity extending a stale tip, a pure loss with no counterparty to sue. And the price of admission for the attacker is bandwidth.

The fix is almost embarrassing in its simplicity: process each transaction announced in a message only once.

Medium severity is the right rating, and it is worth saying why. This does not split the network, does not inflate supply, does not steal a coin. It degrades individual nodes that an attacker chooses to target. The failure mode is delay, not corruption. But delay is exactly the failure mode that matters in a relay race, and the affected versions were 29.0 through 29.5, 30.0 through 30.4, and 31.0 through 31.2. The 29.x series is end of life and will not receive a fix at all. Anyone still running it is exposed by design, which is the least comfortable sentence in the advisory and the one most people will skip.

Three things I could not establish. The $57,655 intraday low comes from a secondary report rather than an exchange tape, so treat it as approximate. My ETF figure covers the fifteen sessions the public daily table still shows, which is the whole of November so far but not a full month. And I have no way to measure how many node operators are still running a 29.x release that will never be patched — nobody publishes a live census by minor version, which means the most important exposure number in this story is also the one nobody can cite.

One hundred and thirty-five days

The disclosure timeline is the real story, and it is not the story the headlines told.

Project Loupe, the automated LLM scanner at Block’s Spiral, reported the issue on 8 July. On 4 August a party calling itself Stutxo reported the same issue independently. That same day Gregory Sanders opened a pull request with a fix. It was merged into master on 6 August. Bitcoin Core 32.1 shipped with it on 27 October. The fix was backported to the 31.x and 30.x branches on 17 November. Versions 30.5 and 31.3 were released, and the CVE published, on 20 November.

Timeline of CVE-2026-117348, from the private report on 8 July to the public advisory on 20 November

Read the dates again and notice what sits between 8 July and 4 August. Twenty-seven days. That is how long the secret held before someone else found it without the benefit of the private report.

This is the empirical fact behind the other change I want to describe, because it destroys the premise that security embargoes rested on for a decade. An embargo is a bet on a gap: the defenders who know about a flaw get time that the attackers do not. The bet pays only if rediscovery is expensive. Twenty-seven days of independent rediscovery is not an expensive bug.

Bitcoin Core changed its security policy on 22 October, and said why. Under the old rules, medium and high severity flaws stayed under embargo until two weeks after the last affected version reached end of life, up to about a year of silence. The new rule lets maintainers flag a bug during triage as one that “can be found with a simple prompt to an LLM,” and those get a fix copied back to every supported version, a point release as soon as possible, and an advisory with details attached. Poinsot, who wrote the change, reported a rise in vulnerability reports for flaws that are easy to rediscover and named the cause: “This is largely due to LLMs drastically reducing the bar to find certain classes of issues or identify covert fixes.”

The scale of what forced that decision is in the numbers. A volunteer group of sixteen developers calling itself the Bitcoin Red Team ran AI models across 390 bitcoin projects in about a day in late July and logged 4,962 findings, 85 of them critical. Within days, a BTCPay Server flaw that had been reported privately was being exploited on live merchant servers, draining the Lightning nodes those merchants used to accept payments. Core Lightning told operators to disconnect after AI-generated reports to its small team turned out to be real.

Not everyone agreed with the new policy. Antoine Riard, a longtime bitcoin security researcher, wanted to keep the option of shipping a fix with no advisory and holding the details case by case. Poinsot refused, on the grounds that telling users a vulnerability exists without saying what it is invites a rushed migration by people who were never at risk. The concession is written into the change: fixes “won’t have as much bake time” before they reach users.

You can argue either side of that trade. What you cannot argue with is the shape of the thing. A security tradeoff was made in the open, by named people, with a named dissenter, in a patch anyone can read, and it is reversible if it turns out to be wrong. Compare it to the process by which the world sets the price of money, which this week produced a 98 percent probability of a rate increase and no reasoning that anyone outside the building can audit.

The case against all of this

I should give the bear case its best form, because this week it earned it.

Nik Bhatia of The Bitcoin Layer has been warning since 22 October, and he is not a permabear by temperament. His framework is liquidity, and it said the dollar was getting strong and tight while everything priced in dollars would have to pay for it. Since his call, the Nasdaq-100 tracking fund QQQ has fallen about 12 percent. On Friday he wrote that high-yield credit spreads were widening again, that Treasury yields were making lower highs as money bought safety, that Korean and Japanese equities had erased their 2026 gains, and that the Brazilian currency was cracking. His conclusion was blunt and uncomfortable: “there is a certain point at which one must admit that the US dollar and US Treasuries serve as a SUPERIOR asset class to bitcoin when times are most troubling.”

This week, they did. That is not a rhetorical point, it is a measured one. Ten-year Treasuries paid you 10 basis points of price appreciation for being frightened. Bitcoin charged you 6 percent for the same emotion. Gold charged you 4.6.

So the strongest version of the bear case is not that bitcoin is a scam. It is that the last eighteen months of the “debasement trade” were a liquidity trade wearing a monetary costume, and that when liquidity withdrew, the costume came off. If that is right, then the four-year-cycle people and the hyperbitcoinization people were making the same error in opposite directions: treating a discount-rate story as a supply story.

There is a second, less comfortable point, and it comes from inside the bitcoin camp. Cory Klippsten of Swan made the honest version of the paper-bitcoin argument this week. The London Gold Pool really did suppress a price, and we know it because the central banks had to keep delivering physical metal until the market closed in March 1968. Futures cannot do that to bitcoin, because a futures contract settles in dollars and the coins never move. But then he concedes the part most maximalists skip: a futures contract “can take a buyer out of the spot market,” and a hedge fund buying coins while shorting a richer future looks, in the positioning data, exactly like someone betting against bitcoin. And the people actually hurt by paper claims were not victims of a conspiracy. They were customers of FTX and Celsius, whose balances were keystrokes.

The same discipline has to apply to the good news. “Easily discoverable” is not a compliment to the code. It is a warning that the class of tool that let sixteen volunteers log 4,962 findings in a day is available to people who will never send a pull request. Nothing about the mechanism that found this bug distinguishes a defender from an attacker except what they do next.

The base layer, priced at thirteen dollars

While the price was collapsing, Block published data on what people actually do with bitcoin at the point of sale: completed payments over $1 at US Square sellers from 1 November 2025 through 15 October 2026.

The median payment was about $13. The average was about $29. More than 42 percent of payments were under $10 and 65.1 percent were under $20. Food and drink businesses took 72.3 percent of them, and within that category coffee shops and fast food accounted for 62.6 percent, with a median of $9.10. Among payments Block could tie to Cash App customers, 44 percent were repeat purchases at the same merchant, and about 86 percent of the gaps between repeat purchases were under 30 days. Most striking: 56.6 percent of payments came from Lightning wallets outside Cash App, and those carried 66 percent of the value.

Read that last number again. A majority of the payments, and two-thirds of the money, arrived from wallets that Square does not own and cannot lock. That is what an open monetary network is supposed to look like, and it is happening at coffee counters for the price of a sandwich.

Two honest caveats. These are Block’s own figures about Block’s own product, published while Block advocates a tax change that would help that product. And Block withholds aggregate volume and payment count, so we know the shape of the distribution and not its size. A median is not a movement.

Meanwhile the governance question that has no price feed got a serious proposal. A developer posted a Node Attestation Protocol to Delving Bitcoin. Node operators publish free-form key-value statements attached to cryptographic proof of control of specific UTXOs. The statements travel peer-to-peer, get aggregated by anyone willing to run an aggregator, and are weighted by coin value. They expire after 26,208 blocks unless re-signed, with a four-week minimum coin age and a 10,000 satoshi floor. It is explicitly non-binding; nobody proposes to gate consensus on a vote. But it is an attempt to answer a question that miner signaling answers badly: what do the people who hold the coins and run the nodes actually think? The author names its weaknesses himself. Attestations reveal which coins you control. Custodians can attest on behalf of owners who never agreed.

That is the work. Unglamorous, public, argued over, correctable.

Two clocks

There are two clocks running on bitcoin, and they keep different time.

The first is set by the credit market. It runs on the discount rate, on the AI capital cycle, on whether December delivers 25 basis points or 50. This week it ended 54 percent below the October 2025 record above $126,000, and it got there in the most ordinary way imaginable: leveraged longs were liquidated, funds redeemed, miners sold, and the price found a new level. Nothing in that sequence required bitcoin to be broken. It required only that people had bought it with money they did not have, at a rate that was about to change.

The second clock is set by the review queue. It runs on pull requests, disclosure policy, and whether a fix reaches the node operator still running an end-of-life release. This week it gained: a deafness bug closed across three branches, an embargo policy rewritten because its premise had expired after twenty-seven days, and a proposal for measuring economic-node sentiment published for anyone to attack.

If you are holding bitcoin because of the first clock, be honest that you are long the Fed and the AI capex cycle, and that the correlation will reassert itself every time liquidity turns. There is no shame in that trade. There is only shame in calling it sound money while checking the price every four hours.

If you are holding it because of the second clock, this week was not a catastrophe. It was the cost of admission, invoiced in the only currency the market accepts. The blocks came. The fix shipped. Nobody had to be asked.

And the practical test is not philosophical. On Saturday morning, before you read anything, which number did you look up first — the price, or whether the network was still producing blocks?

Sources

Write a comment