Decent Newsroom
Sign In

The Option to Reverse

By Roger October 30, 2026
Binance lost $6.3 billion in forty minutes, and only the permissioned part could be clawed back, because stablecoins have a kill switch. Custody, balance sheets, megawatts and benchmarks spent the same week losing their own option to reverse.
The Option to Reverse

The Option to Reverse

At 20:24 UTC on 29 October, wallets labelled as Binance hot wallets began to empty. Forty minutes later the outflow stopped. On-chain tallies compiled by named analysts and first reported by The Block put the total at up to $5.8 billion across Ethereum, Bitcoin and Tron. The receiving addresses had been created for the occasion.

The breakdown that circulated that evening put $3.51 billion on Ethereum and $630 million on Tron. Bitcoin accounted for 18,270 coins, worth roughly $1.44 billion at the time. Bitcoin fell from a 24-hour high near $81,890 to $76,301 as the numbers moved. Across crypto, $1.5 billion of leveraged positions were closed out, and nearly 84 per cent of them were longs, with Binance handling almost half (CoinDesk, 30 October).

By the next morning Binance had a figure of its own. The exchange put the estimated total impact at around $6.3 billion. Roughly $3 billion had already been frozen on-chain with the help of blockchain partners, and the entry point was a compromised laptop belonging to a DevOps engineer. From there the attacker reached a service container and submitted unauthorised withdrawal transactions. Binance said it found no evidence of a private key leak and that the SAFU fund plus company money would cover the losses.

At 09:30 CET on 30 October bitcoin traded at $77,869 (own measurement), about $4,000 below where it sat before the outflow. BNB was the harder hit, down 7.1 per cent.

Almost every number published since is a price on one question: who can undo what, and what the undo costs. Custody answered first, and answered with a kill switch.

The recoverable fraction

Tether’s chief executive said the company froze more than 1.567 billion USDT. Circle’s chief executive said more than $800 million of USDC was frozen in coordination with law enforcement. Changpeng Zhao, still a shareholder, credited Tether with about 971 million and Circle with about 823 million, for roughly 1.79 billion combined. The tallies do not agree with each other, which tells you how fast they were assembled.

Run the arithmetic. The confirmed stablecoin freezes of about $1.79 billion came to 32 per cent of the $5.57 billion outflow floor that TFTC reconstructed from on-chain data. Binance’s own numbers are more flattering: $3 billion frozen against a $6.3 billion impact is 47.6 per cent. Both figures are my division of published numbers, not company statements.

Two ways to count the Binance hack: total outflow and the frozen portion, 29 to 30 October 2026

This is the strongest public case for custodial stablecoins ever assembled in one night. The freeze was fast, it was coordinated with police, and it converted a total loss into a partial one inside a few hours. Anyone who has watched a hacked exchange turn into a multi-year claims process should recognise what that speed bought. They should also notice who paid for it: the people whose balances were denominated in someone else’s liability.

The same mechanism explains the ceiling. Those assets froze because a company controls the contract that issues them. The 18,270 bitcoin and the ether in the attacker’s addresses have no issuer, no support line and no pause function. Per the same on-chain tracking, they were already moving hours later. The recoverable fraction was the permissioned fraction.

A freeze is not a security feature that switches itself on when something is stolen. It switches on when the holder of the minting keys decides that it should, and the reasons available include a hack, a sanctions list and a prosecutor’s letter. Every holder of that asset class watched a third party exercise the power on a Thursday night, at a moment none of them chose.

So the 32 per cent is not a fact about Binance’s security posture. It is a fact about the composition of its float, and about how much of a customer’s balance is an IOU that somebody else can switch off.

Smaller editions of the same night ran in parallel. zkLink Nova stopped producing blocks after an attacker pulled more than $780,000 out of its bridge by injecting a forged batch root through a public relayer (Unchained, 30 October). The Binance attacker, meanwhile, paid for the operation by selling tokens into markets too thin to absorb them: a little-known 2017 token fell 90 per cent (CoinDesk).

The perimeter nobody can inspect

In 2018 Binance created the Secure Asset Fund for Users, funded by a slice of trading fees and marketed as the answer to exactly this night. Eight years on, no audited balance has been published. TFTC’s writeup puts the open question plainly: whether the reserve was sized for a sub-billion-dollar event or a six-billion-dollar one decides whether a solvency conversation is coming.

Binance promised an updated proof of reserves once remediation is complete. That promise is doing more work than it looks. TFTC’s update cites a CoinDesk review of Binance’s own proof-of-reserves data. It found the emptied wallets are the same ones the exchange uses every month to demonstrate that customer assets are backed.

That is the part worth sitting with. The drained addresses were not concealed. They were published, monthly, as evidence of solvency. A proof of reserves is a list of addresses and balances at one moment; it answers whether assets existed at that moment. It does not establish that liabilities are smaller than those assets, and it says nothing about what happens to the same coins during forty minutes of unauthorised withdrawals.

The honest reading of “user funds are safe” is that it is a balance sheet claim rather than a cryptographic one. It may well be true. It is not the kind of claim a customer can check, and the difference between those two things is the entire product category.

Why bitcoin held, and why that is not the same as strength

Bitcoin spent October near $80,000 while every macro desk worried about the AI trade, and Galaxy’s Alex Thorn has a tidy explanation. “None of the people freaking out about AI own Bitcoin,” he said on Unchained. They had “already gotten out of the trade”, so bitcoin was “not in their inventory to sell”.

His evidence is 13F filings. Nine filers cut spot bitcoin ETF positions by 29 to 100 per cent each and shed about 68,000 BTC between them, Goldman Sachs, Millennium and Capula among them. Across all 2,200 filers, combined holdings fell 14 per cent between September 2025 and June 2026. The aggregate moved from the equivalent of 355,000 BTC to 307,000. That net decline of 48,000 BTC is smaller than the 68,000 the nine sellers gave up on their own. The remaining 2,191 filers were therefore net buyers of about 20,000 BTC (own arithmetic).

The steel-man version is that bitcoin’s marginal holder is now a wealth manager with a quarterly rebalancing calendar rather than a macro fund with a stop-loss. Patient capital produces shallow drawdowns. The unflattering version is that low volatility with no marginal buyer looks identical to resilience on a chart. Both descriptions fit the same month of data, and the filings due on 16 November will separate them.

There is also an inconvenient detail in the last 40 minutes. Bitcoin fell 6.8 per cent from its intraday high to its low, which is roughly the same shock BNB took. The resilience claim is about the month, not about the hour, and the hour looked like every other risk asset.

The balance sheet that became a credit book

Strategy reported the largest operating income in its history for the third quarter, $20.91 billion (Bitcoin Magazine). Net income came to $19.53 billion, or $44.75 per diluted share, against $2.79 billion a year earlier. Most of that is a mark-to-market gain on the bitcoin stack, which is what happens when a treasury is marked quarterly and the price rose.

The company held 848,517 bitcoin at an average cost of $75,446, or $64.02 billion (own arithmetic). At my measured $77,869 the position is worth $66.07 billion, a paper gain of $2.06 billion, or 3.2 per cent. Five days earlier, at $81,200, the same coins were booked at $68.9 billion.

The per-share picture moves the other way. Holdings rose 26 per cent year to date while bitcoin per share, the metric management uses as its own yardstick, fell 3.7 per cent. Divide one by the other and the share count grew about 30.8 per cent over the same period (own arithmetic). Total revenue fell 3 per cent to $125.4 million and missed estimates, and the company sold roughly $431.8 million of bitcoin during the year to fund preferred dividends and buybacks.

That last line is the tell. The preferred stock is not a side quest. Strategy repurchased 20,980,508 shares of STRC for $2.04 billion at an average of $97.39, a discount to the $100 stated amount, and about $956.7 million remains under the authorisation. It kept the STRC dividend at 12.00 per cent to support trading near par. Shareholders approved daily dividends on 28 October: STRC pays daily from 1 November, with STRF, STRK and STRD following on 1 January 2027. Dollar reserves of $5.07 billion are said to cover 3.4 years of dividend and interest obligations.

A company buying its own 12 per cent preferred below par is saying two things at once. The buyback is accretive to common holders, and the market spent the quarter doubting that the stated amount would be honoured. Both can be true. The second one is what costs money, and it is why the reserves disclosure now matters more than the bitcoin count.

The megawatt that cannot come back

The quietest reversal in the whole week is happening in power contracts. CoinShares data reported by TFTC puts AI and HPC workloads at roughly $1.5 million of revenue per megawatt per year. Bitcoin mining earns about $500,000 on the same megawatt, a threefold spread consistent across two quarterly reports. IREN became the first large listed miner where AI cloud revenue ($70.5 million) beat mining revenue ($66.7 million) in a single quarter, a 5.7 per cent edge.

The headline about the sector is a backlog above $100 billion in disclosed AI and HPC contracts. The number attached to that headline is smaller. Annualised billing revenue is roughly $1.1 billion, against about 550 megawatts actually live out of more than 4 gigawatts under contract, or 13.8 per cent (own arithmetic).

Two sanity checks make that gap less mysterious than it looks and more worrying. Divide $100 billion by 4,000 megawatts at $1.5 million per megawatt-year. The implied average term is 16.7 years, which is what a long-duration lease book looks like rather than a revenue line. Divide $1.1 billion by 550 live megawatts and the implied revenue is $2.0 million per megawatt. That sits above the planning figure and suggests the live sites carry the richest contracts (both own arithmetic).

Miners sold the megawatt: revenue per megawatt for mining versus AI tenancy, and contracted versus live capacity

Here is what my own node measures this morning, at block 973,794. Difficulty stands at 148.36 trillion, network hashrate at 924 exahashes per second, and the mempool holds 31,886 transactions in 8.7 megabytes. Across the last fourteen blocks, fees averaged 0.0203 BTC against a 3.125 BTC subsidy, so fees carried 0.65 per cent of the block reward. At $77,869 the annual security budget comes to $12.87 billion, of which 99.35 per cent is new issuance. Hashprice works out at $38.17 per petahash per day.

The annual security budget today and after the 2028 halving, with the measured fee contribution

Two numbers put the megawatt question in scale. Assume 18 joules per terahash, the efficiency of current-generation machines; that is an assumption, not a measurement. On that basis 4 gigawatts of capacity is about 222 exahashes per second, roughly 24 per cent of the 924 I measure. And 76,206 blocks remain until the 2028 halving, about 529 days, after which the subsidy halves and the subsidy-funded budget at today’s price falls to $6.40 billion (own arithmetic).

AI conversion capex runs $10 to $15 million per megawatt, against $1 to $1.5 million for mining infrastructure. A fully leased AI data centre recently transacted near $27 million per megawatt; energised but unleased miner capacity goes for below $3 million. Money at that size does not come back quickly. The leases are long, the concrete is poured, and the price that would pull those megawatts back to hashing sits far above today’s.

The energy that makes it bind

Cheap power and cheap money were the two inputs both industries were built on, and on 29 October the European Central Bank declined to make the second one cheaper. It held the deposit facility at 2.50 per cent, the main refinancing rate at 2.65 and the marginal lending facility at 2.90. The statement gives the reason: energy prices “have continued to rise”, with “increased magnitude and duration of the energy shock” likely to push inflation up and growth down. It added that financing conditions have tightened since the previous meeting.

The market side of that story is visible in the same morning’s prices (own measurement). WTI crude trades at $103.65 and gold at $3,999.70. The US ten-year yields 4.931 per cent against a 2.50 per cent ECB deposit rate, a spread of 243 basis points. A world where energy is expensive and money is not getting cheaper is precisely the world in which a threefold revenue gap per megawatt becomes impossible to ignore.

Two claims you can actually check

Set against all of that, a post on Delving Bitcoin this week is almost restful. OP_MUL has been disabled since 2010, so multiplication inside Tapscript has to be built from addition and shifts. Every intermediate that a later opcode reads must still fit in a four-byte CScriptNum. The tightest symmetric bound that survives that rule is the floor of the square root of 2^31 minus 1, which is 46,340. Squaring it gives 2,147,395,600, leaving a margin of 88,047 below the ceiling.

The author released a 371-byte unrolled multiplication core that stays inside that bound, with 13 active opcodes and a peak stack depth of 21 items. The witness supplies the bits of the multiplier, and the script reconstructs it and verifies the reconstruction. A 275-byte variant drops redundant guards when you are willing to rely on MINIMALIF, saving 96 weight units. The Gist carries a Rocq 9.3.0 proof with six closed theorems and no axioms. A single-file Rust auditor sweeps 46,341 rows in 54 milliseconds, certifying 2,147,488,281 pairs, and kills 455 mutants. The witness was tested against Bitcoin Core 31.1.0 in regtest through testmempoolaccept.

The post ends with questions rather than conclusions, including whether paying 96 weight units for guards that do not depend on SigVersion is worth it. That is what a checkable claim looks like: a size, a proof, a cost, and a list of the things it does not do.

The same instinct showed up on the AI side, in a smaller and messier way. A benchmark called DecoyCount tests whether a model is counting an object or falling for a reflection, a shadow or a substituted duplicate. It separates the seven ways that test can fail. Across 1,660 items the best frontier model, GPT-6 Astra, scored 43.17 per cent. Gemini-3.8-flash managed 32.71 per cent. Qwen3.8-max reached 28.31 and Claude-Fable-5.1 25.22. Elorian’s Glimpse, the model the benchmark’s co-authors built, reported 50.48 per cent, in a single run without a tool harness, against three-run averages for the others.

The paper’s own conclusion is the sentence to keep: “stronger general reasoning does not by itself generalize to aligned visual reasoning”. A score below 50 per cent on counting is not a headline about any one lab. It is a statement about how thin the evidence under a leaderboard can be. The researchers published the failure modes alongside the number, which is the part that makes the number worth anything.

What would change my mind

Four things, each with an address.

If the SAFU fund pays every user in full and quickly, without haircuts or withdrawal gates on a KYC pretext, the custodial model performed as advertised. If Tether and Circle then return the frozen coins to identified owners rather than merely immobilising them, it performed better than any exchange hack in this asset class has. That outcome would be the first of its kind, and it would weaken the verification argument for anyone who keeps trading balances on an exchange.

If the Q3 filings show billed megawatts climbing toward the 4 gigawatts already under contract, the pivot is a revenue story and my irreversibility point is early. Watch billing, not signatures.

If fees stay under 1 per cent of the block reward through the 2028 halving, the security budget question stops being a talking point. It becomes arithmetic that price has to answer. If instead fees rise above 5 per cent while hashrate holds, the network found the second funding leg without needing a crisis.

And if the 13Fs due on 16 November show those nine sellers buying back, then the line about AI panic and bitcoin ownership described one quarter rather than a structure.

Two things are not verifiable from outside, and I will not assert them. Binance’s $6.3 billion is the exchange’s own estimate, and no independent reconstruction of the full figure exists. The SAFU balance is unknown outside the company. The 550 megawatts figure is CoinShares’, and whether it measures facility power or tenant IT load changes every per-megawatt ratio in this piece.

The thing that kept its value

Custody, capital and megawatts all gave up their option to reverse this week, and each of them sold it at a price somebody else set. Binance’s customers held claims that were made whole by a balance sheet and by two companies with a freeze button. Strategy’s shareholders hold a paper gain that is now smaller than it was on 23 October, financed by preferred stock the company itself is buying back below par. The miners sold the option on their own sites for sixteen-year leases.

The only item on the board that did not need anyone’s permission to stay true was arithmetic. A 371-byte script either stays inside 46,340 or it does not, and the difference is 88,047, checkable in 54 milliseconds by anyone who wants to bother. That is a strange thing to call the strongest asset of the week, until you notice what everything else cost.

Sources

Write a comment