Nineteen Cents
- Nineteen Cents
- The contract that saw it first
- The wrapper that was not a share
- What a signature authorizes
- The largest fund in the world had its worst day
- Strategy pays the preferred first
- Custody moved, it did not shrink
- What the machines are actually doing
- The steel-man case against my own argument
- Leverage is the only thing that pays
- The honest part: what I could not check
- What to watch
- The last thought
- Sources
Nineteen Cents
At 11:47 on Thursday morning in New York, a perpetual futures contract on the crypto exchange OKX printed $111.31. One minute later the first trade in Anthropic’s stock printed on Nasdaq at $111.50.
Nineteen cents. On a company that had never had a public price, in an offering that raised $81.3 billion and valued the firm at roughly $2 trillion fully diluted. The regulated venue, with its opening auction, its designated market makers and half a century of price discovery, produced a number that a levered contract settled in dollar tokens had already found, to within two tenths of one percent, two hours earlier.
This is being reported as a curiosity about crypto markets. It is better read as a lesson about what a market price is. A price is not a fact about a company. It is the current best guess of the marginal buyer with money at risk, and on Thursday the marginal buyer with money at risk was sitting in front of a screen in Asia, not on the floor of an exchange in New Jersey.
That same week, three other claims on assets were repriced, and they were repriced in the opposite direction. The largest bitcoin fund in the world had its worst day since it launched. The largest corporate holder of bitcoin fell for the sixth session in a row. And the token that ordinary people could actually buy to get exposure to Anthropic had to be refunded in full, because the allocation the issuer received was, in its own words, “not ideal relative to total user commitments.”
Four instruments, one question underneath all of them: when you buy a thing, what exactly do you own, and who can take it from you without asking?
The contract that saw it first
The mechanics are worth spelling out because they explain why the derivative beat the stock.
Anthropic priced at $83 on Wednesday night and sold 980 million Class A shares. The roadshow range had been $70 to $78; the book came in above it. On Thursday, before the Nasdaq opening auction, two venues offered perpetual futures on the ticker: OKX, and trade.xyz on Hyperliquid. These are contracts with no expiry, no share delivery, settled in crypto, margined continuously. At 9:30 a.m. New York time they traded at $103.75 and $103.50, about seven percent below where the stock eventually opened. By 11:47 they had converged to $111.31 and $111.00.
Then the stock printed at $111.50, ran to $114.99 within minutes, slid to $100.43 by 12:29 p.m., and closed at $97.17, up 17 percent from the offer price and 13 percent below its own opening print. After hours it drifted to $92.60.
Two things follow. First, the perp market’s error on the opening price was 0.2 percent, against an average miss of 25 percent on 2026’s earlier listings measured by Unchained in October. That is not a rounding improvement; it is a market learning to do an opening auction without an auction. Second, and more important: while the stock had no price at all, the perp had one. At 9:30, an hour and a half before the opening auction, it was already quoting within seven percent of the print that would eventually clear. A trader with a view on Anthropic did not need Nasdaq to express it, and Nasdaq was about to price against a number the crypto venues had published first.
Note what the perp is not. It is not a share. It confers nothing: no vote, no dividend, no claim in bankruptcy, no information rights. It is a bet on a number, and it does the job of price discovery better than the instrument that carries the ownership rights, because it has no settlement friction. That is the trade-off the whole week was about. The instrument with the least ownership did the best job of pricing. The instruments with the most nominal ownership did the worst job of protecting their holders.
The wrapper that was not a share
While the perps were converging, Binance Wallet was running a subscription for ANTx, a token issued under the xStocks label by Backed Assets (JE) Limited, tracking Anthropic’s stock without conveying the shares. Users committed the USDC stablecoin before the IPO price was known, against an indicative range of 70 to 78 USDC per token, plus a five percent xStocks fee. The 500 million USDC cap was hit on Wednesday.
On Thursday Binance Wallet announced that the IPO allocation it received was “not ideal relative to total user commitments.” It returned all locked USDC, airdropped the allocated tokens to participants at no cost in proportion to what each had committed, and added one million dollars of its own bStocks Anthropic token to everyone who applied.
Read that sequence again. The product sold out. The issuer then discovered it could not deliver what had been sold, so it gave the money back and handed out the asset for free. That is not a functioning primary market. That is a queue that dissolved when it reached the front.
The fine print had said so in advance. Ondo’s release states that its ANTon token is “not issued, sponsored, or endorsed by Anthropic.” OKX’s listing page states that its tokenized stocks do not “confer shareholder rights such as voting.” And Anthropic itself, before the listing, warned on its support site that sales of its private stock without board approval were void, naming tokenized securities among the ways such sales were being offered.
So the buyer of the wrapper was holding a claim on an issuer’s ability to obtain the underlying, sold at a five percent premium to an indicative price that turned out to be wrong, on an asset whose actual issuer had publicly declared the whole construction void. Every one of those facts was disclosed. None of them was in the marketing.
This is the same failure the ETF wrapper exhibits in a slower, more respectable form. But before we get to that, there is a paper published this week that describes the mechanism exactly, in a completely different domain, and it is the best thing I read all week.
What a signature authorizes
arXiv 2611.24989, “Signed, Valid, and Unauthorized,” looks at agent commerce protocols, the layer where a software agent pays with stablecoins on behalf of a human. The buyer approves a payee, an asset, a spending cap and a time window. The wallet then signs an object, and the effects of that object depend on the contracts that consume it.
The authors define something they call authority closure: the set of all transfer sequences that holding the signed object can enable. A wallet object respects the mandate only if every sequence in that set meets the buyer’s constraints. Then they test it. Across 2,130 generated objects and 62,979 execution programs, 265 objects with an observed mandate violation passed both a field-and-expiry check and an address allow-list. Both of the checks a careful engineer would build. Both of them blind, because they inspect what the object displays rather than what it authorizes.
Their checker, Closure, rejected all 1,680 objects with an observed violation and accepted the remaining 450, with a median decision time of at most 12 microseconds per family.
Taylor Monahan of MetaMask, speaking on Unchained on 18 November, gave the security industry’s version of the same gap: “the audit gets us 90 percent,” formal verification adds another five, and “you still have 5 percent that it might just completely collapse.” Her answer is a circuit breaker: a rule that says that under no circumstances should $200 million ever leave this contract, and an automatic shutdown if it does. That is authority closure implemented as a tripwire, and it is the one control that does not depend on predicting the attack.
Hold that next to the ANTx subscription. The buyer saw a ticker, a price range, a five percent fee, a blockchain, a brand. All the displayed fields matched a mandate that said “buy Anthropic exposure.” The set of things the object actually authorized did not include owning a share of Anthropic, did not include a claim on the allocation, and did not include any right that survives the issuer’s discretion. The fields matched. The authority did not.
This is not a crypto problem or an AI problem. It is the general problem of delegation: whenever you sign something whose effects are determined downstream by someone else’s contract, your approval of the displayed terms is not the same as approval of the reachable outcomes. For a human clicking through a brokerage app the gap is usually small and legal. For an autonomous agent holding a key, it is the entire risk surface, and the paper’s contribution is to show that it can be closed cheaply, 12 microseconds, if anyone bothers to compute the closure instead of checking the fields.
Anyone shipping an agent that can move money should read it. Not because the attack is exotic, but because the fix is trivial and the failure mode is silent.
The largest fund in the world had its worst day
Now the other side of the week, where the claims were older, larger and better lawyered.
On 19 November, BlackRock’s IBIT recorded $541.1 million in net outflows, the largest single-day redemption since the fund launched in January 2024. The whole US spot complex shed $582.3 million that session, so IBIT alone was 93 percent of it. Fidelity’s FBTC lost $45.9 million, ARK’s ARKB $31.3 million. Morgan Stanley’s MSBT took in $49.6 million. Everything else was red.

My own tally of Farside Investors’ daily table, read on the morning of 20 November, gives fourteen sessions from 2 to 19 November with exactly one positive day (4 November, plus $74.6 million) and a cumulative net outflow of $3.49 billion. (Own calculation from the Farside table; individual daily figures are Farside’s.)
The framing that this is “record outflows, therefore bitcoin is falling” is true and insufficient. The interesting part is the plumbing.
An ETF share is a claim on a trust that holds bitcoin. When a large holder redeems, the authorized participant delivers shares and receives value. That value can be settled in kind (actual bitcoin transferred out) or in cash, in which case the AP has to sell bitcoin to fund it. The choice is operational, and it is invisible to the shareholder. A self-custodied coin has no equivalent mechanism. There is no AP, no creation basket, no redemption queue, no trustee who can be instructed to liquidate. The only seller of a coin in cold storage is its owner, at a price he chooses.
IBIT’s own numbers make the point sharper than any argument. Its product page put year-to-date NAV total return at −28.53 percent through 18 November, with net assets around $48.36 billion. Its 10-Q for the quarter ending 30 September showed 801,033 bitcoin held. Whatever the pitch was in January 2024, a holder sitting on a 28 percent loss who redeems is not patient capital, whatever the wrapper’s brochure said. Patient capital adds at minus 28. Momentum capital leaves. The wrapper did not convert one into the other; it only made the exit faster, because exiting an ETF takes a click, while exiting self-custody takes a decision, a fee and an address.
Bastiat’s formulation fits: the outflow is what is seen. What is not seen is the holder who did not have to sell because nobody could force him to.
Strategy pays the preferred first
The second repricing was the leveraged version of the same claim, and it is more instructive because the balance sheet is public.
Strategy closed at $86.13 on 19 November, down 4.6 percent on the day, its sixth consecutive decline. That is 37 percent below the $136.02 close of 19 October, and 43 percent below the $151.95 at which it ended 2025. Bitcoin over the same stretch went from $81,938 to $61,987, a decline of 24 percent. Same direction, roughly one and a half times the amplitude.

The arithmetic on the treasury is now straightforward. In its 8-K of 16 November, Strategy reported buying 485 bitcoin in the week ending 15 November for $33.4 million, an average of $68,887. Total holdings: 850,849 bitcoin, acquired for $64.20 billion, an average of $75,452. At Thursday’s price the stack was worth roughly $52.8 billion, an unrealised loss of about $11 billion, and bitcoin was trading 18 percent below the company’s average cost.
Here is the part that deserves more attention than the headline loss. In the same filing, Strategy reported repurchasing 2,282,510 shares of its STRC preferred for $228.1 million, nearly seven times what it spent on bitcoin that week, funded with $244.7 million raised by selling common stock. It still has $332.2 million left under the preferred repurchase programme and $1.0 billion under a common buyback authorisation. Michael Saylor said on X that the buybacks add liquidity, lower dividend obligations and strengthen the credit, and that the company intends to keep buying below $100.
That is the honest description of what the flywheel has become. A company that for three years converted equity and convertible debt into bitcoin is now converting equity into the retirement of a preferred instrument whose dividend it owes in dollars. The bitcoin stack is not being increased; it is being defended. Nothing in that is fraudulent or even unreasonable: a 4.9 billion dollar reserve of dollars exists precisely to service those obligations. But it is a different business from the one the multiple was built on. The marginal buyer of MSTR today is not buying a bitcoin proxy at a premium. He is buying the residual claim on a stack that is under water, behind a preferred that just consumed seven times the week’s bitcoin budget.
In Hong Kong, someone is monetising exactly that. The CSOP 2x Short MSTR product, ticker 07399, rose about 19.8 percent in late trading on 19 November and is up more than 120 percent month to date. Turnover of HK$4.47 million, roughly $575,000, is small. But an inverse product with that month argues that at least one institution’s risk desk has concluded the premium compression is a trend rather than a wobble.
Custody moved, it did not shrink
The third repricing was quieter and, for anyone holding bitcoin inside a wrapper, the one to read twice.
VanEck served written termination notices to Gemini on 16 November covering custody and clearing for all three of its crypto funds (HODL, ETHV and VSOL), effective on or about 16 February 2027, per three 8-K filings submitted on 19 November. No reason was given in any of them. Coinbase Custody remains as custodian for all three. For HODL, Gemini had been the primary bitcoin custodian with Coinbase listed as additional; after February the hierarchy is a single counterparty. For VSOL, Gemini’s staking addendum terminates with custody and no replacement staking provider has been named.
The consolidation is not obviously more dangerous than the arrangement it replaces. It is, however, the removal of redundancy, and it happened three weeks before year-end with roughly ninety days’ notice and no explanation, which is what a risk decision looks like when the lawyers have advised against describing it.
Sit that next to the quantum guidance published by IOSCO this week. The global standard setter for securities regulators told firms not to wait for a cryptographically relevant quantum computer before planning migration, put conservative estimates at 2030 to 2035, and reported that some participants argued for a five-year horizon. It noted that one standardized post-quantum signature scheme, ML-DSA-44, produces signatures of about 2,420 bytes against the 64 bytes of the elliptic-curve signatures used today, roughly 35 to 40 times larger, and that a quantum-resistant Solana testnet saw throughput fall about 90 percent. Migration, in its words, “cannot be accelerated at the last minute.”
For a self-custody holder, a signature-scheme migration is a wallet software update and a transaction he signs. For a shareholder in a fund, it is the issuer’s project, on the issuer’s schedule, with the issuer’s counterparties, and his only lever is to sell. Matt Hougan of Bitwise called quantum “mostly a migration risk, not a meltdown risk” on Unchained on 9 November, except for Satoshi-era and pre-2015 coins whose public keys are already on chain. He is right about the risk and the framing understates the asymmetry: the people who can migrate are the ones who hold keys.
What the machines are actually doing
The AI side of the week produced a paper that belongs in this argument because it prices the thing everyone is financing.
arXiv 2611.23375, “Where the Tokens Go,” treats a coding agent thermodynamically. It counts the tokens an agent generates as cost and divides them into work and heat, with the test suite playing the role of Maxwell’s demon, judging each change. The experiment is an analogue of pulling a single molecule open and letting it fold back: an agent makes a change to a repository and then undoes it, many times, counting every token. About 80 percent of the tokens in a successful run are generated in model calls that leave the requested change exactly where it was. Those are heat. The free energy of a change, meaning what the work actually costs, is estimated at 14 to 22 percent of the mean cost of a successful run. And coding costs 1.1 to 1.5 times as much as uncoding.
Eighty percent heat. That is a measurement of waste inside a single agent loop, and it is the most useful cost number published this year, because it is the only one that separates “the model computed something” from “the code changed.”
Now put the FT’s arithmetic next to it. Per the Unhedged column summarised in Latent Space’s AINews on 20 November: a 20 percent return on roughly one trillion dollars of compute commitments, at a 40 percent operating margin and a 17 percent tax rate, requires about $602 billion of annual revenue.

Alphabet’s annual revenue is about $471 billion. Apple’s is about $478 billion. All 76 software and IT companies in the S&P 1500 together earned $908 billion. The requirement is two thirds of that entire pool, from one company, to service one cohort’s compute commitments at a merely adequate return.
That is the capital-theory version of the week. Capital was sunk into long-lived, immovable assets, data centres with fifteen-year lives, on the basis of a price signal that was itself set by the cost of money and by an equity market willing to pay $97 for a share that was offered at $83 the night before. The asset cannot move. If the revenue arrives, everyone is a genius. If it arrives at 60 percent of the required level, the loss is not the difference between two share prices. It is the difference between what the building can earn and what the paper demands, which is the definition of malinvestment and has nothing to do with anyone’s intentions.
Anthropic, for what it is worth, looks like the least levered participant in this. It raised about $140 billion in total and still holds about $132 billion, which implies a cumulative burn under $10 billion to reach this point. Tanay Jaipuria of Wing VC, on Latent Space, called that the S-1’s biggest positive surprise. AMD put in $2.5 billion alongside the IPO; Google owns 13.7 percent; the Class F shares carry 50.1 percent of voting power; and roughly a trillion dollars of shares come out of lockup over the next year. Dario Amodei used the occasion to commit to pacing training and deployment “taking as long as is necessary to have confidence that they are safe and will remain under human control,” which is a strange sentence to say at the moment your stock begins trading, and a good one.
The steel-man case against my own argument
The obvious objection to all of this is that I am pattern-matching a coincidence. Bitcoin fell, an AI company listed, therefore capital rotated. Cory Klippsten, who runs Swan and has no incentive to be neutral, argued the opposite on 19 November: the SpaceX IPO in June produced far more noise among his clients and contacts than Anthropic did, and this week’s damage tracked Nvidia, not the offering. He would put “more chips” on a general tech freakout than on rotation. That is a serious point and I think it is more right than wrong as a causal claim about Thursday specifically.
The counter-evidence is not a story, it is a schedule. The ETF outflows began well before Anthropic priced: fourteen sessions, one positive day, $3.49 billion net out over the period in which the IPO book was being built. Correlations do not prove flows, but bitcoin’s 40-day correlation with the Nasdaq 100 sits near 0.69 on Bloomberg’s data, which is what a risk asset with institutional ownership looks like, not what an uncorrelated monetary hedge looks like. And the rate market was not offering relief: the two-year Treasury yield closed at 4.80 percent on 18 November against a policy range whose midpoint is 3.875 percent, which is a market pricing more tightening rather than less (FRED, read 20 November). Both things can be true: Nvidia broke the tech tape, and the AI complex is where the marginal dollar went. Rotation is not the cause of the first tick down; it is the reason the bid did not come back.
Franklin Crypto’s Seth Ginns offered the friendlier version on Bits + Bips: a valuation around $1.4 trillion looked like “a fair price,” he expected “a nice pop,” and the listing was “probably a net positive to crypto” because some of the money investors pooled would come back after the pop. He was right about the pop. Whether the money comes back is the open question, and the honest answer is that nobody knows, including him, which is why he prefaced it by saying he didn’t have a strong view on Anthropic.
Leverage is the only thing that pays
One more paper, because it explains why the four instruments behaved differently and it is not about markets at all.
arXiv 2611.23322, “Human-AI Mutualism Is Unstable under Capability Asymmetry,” makes an argument that generalises past AI. Bargaining guarantees each side only what it would get if bargaining failed. Humans stay as well off as they were before AI, absent alignment, only while they retain leverage: either the AI depends on something they supply, or they can impose costs large relative to what the AI can produce. Once neither holds, alignment is the only thing left, and if the other side’s goals are insatiable, the required alignment approaches perfection as capability grows.
Strip the AI out and it is a statement about every claim in this piece. The ANTx buyer had no leverage: no vote, no allocation guarantee, no legal claim, nothing to withhold. The IBIT holder has one lever, the sell button. The MSTR holder is behind a preferred and a board. The self-custody holder has the only kind of leverage that does not require the counterparty’s cooperation: the ability to be somewhere else, with the asset, without asking.
That is what a key is. Not a symbol of ideology, not a hedge against a spreadsheet, but a structural position in a bargaining problem. Everything else in this week’s news was somebody negotiating from the weak side of one.
The honest part: what I could not check
What I cannot show: that a single dollar left IBIT to buy ANT. Flows are not traceable that way and anyone who claims otherwise is inventing a ledger. What I can show is that in the same week, the instruments built to give people exposure to bitcoin without ownership behaved exactly as exposure behaves when the price falls: redemptions, an inverse product up 120 percent, a preferred eating the acquisition budget. The thing itself did what it always does, which is nothing. Bitcoin did not file an 8-K. It did not terminate a custodian. It did not refund a subscription. It went from $81,938 to $61,987 and every coin stayed exactly where its owner had put it.
One more admission. The perps being within nineteen cents is impressive, and it is also a warning. The same mechanism that discovers price accurately for a listed stock will discover it for anything with a payoff, at leverage, around the clock, with no circuit breaker and no exchange to appeal to. The venue that knows the answer first is not necessarily the venue where you can survive being wrong.
What to watch
Three falsifiable markers, all from the reporting above, all checkable without me: whether IBIT stays in net outflow beyond five consecutive trading sessions; whether MSTR’s premium to net asset value fails to re-expand above 1.5 times; and whether bitcoin reclaims $70,000. If all three flip, this week was a flush. If two of three persist, the claims are in a durable drawdown that the underlying is not.
Add two dates. February 2027, when VanEck’s Gemini termination takes effect and HODL’s custody becomes a single counterparty. Watch for an 8-K explaining why, and for whether Gemini picks up a comparable mandate elsewhere. And the lockup calendar, roughly a trillion dollars of Anthropic stock freeing up over twelve months, which is the largest scheduled supply event in the story and the one nobody is pricing in the perp.
The last thought
The week’s lesson is not that bitcoin fell or that AI is a bubble. Both may be true and neither is new. The lesson is that “exposure” and “ownership” became different products, sold under the same name, to the same people, in the same year, and when the price moved, only one of them let you decide what happened next.
Anthropic’s buyers found out that a token tracking a stock is not the stock, and the refund was the kindest possible way to learn it. IBIT’s redeemers found out that a share in a trust is a claim on a trustee’s operations. Strategy’s shareholders found out that a bitcoin treasury with a dollar dividend attached is a credit instrument wearing an equity costume. VanEck’s holders found out that redundancy is not a right. In each case the displayed field matched and the authority did not.
The holder of a key found out nothing new. That is the whole point of holding it. Nineteen cents was the market telling us where price is made. The other three stories were the market telling us where claims end.
Sources
- Unchained — Crypto Traders Had Anthropic Within 50 Cents of Its $111.50 Nasdaq Open
- Latent Space — AINews: Anthropic successfully completes largest IPO in history, valued at $1.5T
- TFTC — IBIT Posts Record $541M Single-Day Outflow as Bitcoin ETF Complex Sheds $582M
- Farside Investors — Bitcoin ETF flows
- TFTC — VanEck Drops Gemini From All Three Crypto ETFs, Coinbase Takes Sole Custody
- Bitcoin Magazine — Strategy (MSTR) Shares Tumble for Sixth Straight Day to $86 as Bitcoin Slides
- Unchained — Global Securities Regulators Tell Financial Firms to Plan for a Surprise Quantum Zero-Day
- arXiv — Signed, Valid, and Unauthorized: Authority Closure at the Trust Boundary Between Agent Commerce Protocols and Crypto Wallets
- arXiv — Where the Tokens Go: Heat, Entropy and Error in Agentic Coding
- arXiv — Human-AI Mutualism Is Unstable under Capability Asymmetry
- Cory Klippsten — Two Down, One to Go: What the AI IPOs Mean for Bitcoin
- Unchained — Taylor Monahan Says Crypto Teams Need Monitoring and Circuit Breakers, Not More Audits
- Federal Reserve Bank of St. Louis (FRED) — 2-Year Treasury constant maturity
- Binance — BTCUSDT daily klines
- stockanalysis.com — Strategy (MSTR) historical prices
Write a comment