Decent Newsroom
Sign In

The Moat Is One Month Wide

By Roger October 25, 2026
A twelve per cent perpetual dividend against an asset that pays nothing. A capability lead one month wide. Gold at $4,008 against bitcoin at $80,300. Three margins were repriced in the same week, and they are the same margin.
The Moat Is One Month Wide

The Moat Is One Month Wide

A perpetual preferred share paying twelve per cent against an asset that pays nothing. A one-month capability gap behind a frontier that costs billions a year to hold. A gold price of $4,008 and a bitcoin price of $80,300. Three margins were repriced in the same week, and they are the same margin.

At 11:20 a.m. Eastern on Saturday, bitcoin printed $79,280. It was the weakest level in more than a month, and it lasted about ninety minutes.

What happened around it is more informative than the print. In the fifteen minutes from 10:15, $208.21 million of long positions were liquidated as bitcoin slipped from roughly $80,640 to $80,100. That is two thirds of one per cent. Another $131.17 million went in the fifteen minutes from 11:15. By one o’clock, of the $436.44 million liquidated in four hours, $404.15 million had been longs, according to CoinGlass data published by The Block. Long liquidations between eight in the morning and noon came to $401.91 million, the heaviest four-hour stretch since the night of 14 October, when $505.57 million was wiped out.

Over twenty-four hours the total reached $815.26 million across 201,012 accounts, up 218 per cent on the previous day. Open interest in crypto futures fell 2.2 per cent to $140.10 billion. Spot bitcoin ETFs lost $773 million on Thursday and Friday, ending three days of inflows. Arthur Hayes, who is rarely quiet about his book, wrote overnight that he had sold it: “I dumped it all.”

A two-thirds-of-a-percent move does not liquidate four hundred million dollars unless the structure underneath is levered far past the move it is designed to survive. That is the first of three margins this week, and it is the least interesting of them, because it resets by itself. The other two do not.

Where the $436 million of Saturday liquidations went, by fifteen-minute window

What the price series actually says

The headline, bitcoin below $80,000, reads like a break. The daily closes say something narrower and more useful.

Bitcoin’s high of the past twelve months was $124,658 on 6 October 2025, a year ago this month. It closed at $86,637 on 20 November, $66,461 on 18 February, and $58,625 on 30 June. From that June low it recovered to $86,620 on 21 September, then gave ground: $83,624 at the end of September, $81,682 on 14 October, $80,439 on the last daily close. Measured from the October 2025 peak, Saturday’s level is 35.5 per cent lower. Measured from the June trough, it is 37.2 per cent higher. Measured from the first close of this year, it is down 9.5 per cent.

Every one of those figures is my own reading of 400 daily closes, pulled on 25 October. They are worth stating precisely because the story they tell is not the story the headline tells. Bitcoin has been in a bear market for twelve months. It bottomed in June, rallied 48 per cent into September, and has spent October surrendering most of that rally. Saturday was not the beginning of a decline. It was the failure of a recovery, and it happened to land on a weekend when leverage was thick.

That distinction matters for what follows. A market breaking to new lows is repricing. A market failing a recovery is being told that the bid is not there at the price the holders need. The second is a statement about the buyer, not the asset.

Bitcoin daily closes over 400 sessions, October 2025 to October 2026

The dividend that comes out of the asset

Here is a capital structure that has quietly become one of the largest sources of bitcoin demand, and that almost nobody outside the trade can describe in a sentence.

Since 2025, companies have sold perpetual preferred stock and used the proceeds to buy bitcoin. Fidelity’s research analyst Zack Wainwright counted 135,979 coins funded this way through 5 October: more than half of one per cent of the maximum supply of 21 million, from a market Fidelity sizes at roughly $12 billion. Strategy alone has five classes of the stuff. Its first, STRK, debuted in February 2025.

Its design is the whole argument, so the numbers only make sense after the mechanism. A perpetual preferred has no maturity date. That is its selling point: no refinancing cliff, no rollover risk, no need to find a new buyer at a fixed date to repay the old one. What it has instead is a coupon that never stops. Strategy’s STRC pays a variable dividend currently around 12 per cent, according to CoinShares’ head of research James Butterfill. A yield investor buys it, comparing a perpetual claim against credit, and 12 per cent is a number that competes.

Now do the arithmetic that the structure implies. The asset pays nothing. The claim pays 12 per cent, in cash, forever, and ranks ahead of the common shareholder. For the common shareholder to be made whole, the asset must compound faster than the cost of the claim against it. Nobody knows whether bitcoin will compound faster than that. What is certain is that the instrument only works if it does, every year, rather than on average.

Two things follow, and both showed up in the reporting this week.

First, the machine is reflexive in a way that ETFs are not. Mark Palmer of StoneX put it plainly: continued preferred-funded buying “depends on the securities trading at or near par.” Above par, issuing more is accretive to bitcoin per share; below par, each new sale hands the buyer more claim than it hands the company asset, and the rational move becomes buying the paper back instead of issuing it. STRC traded around $99.40 in the week, close enough to par that nothing is broken, and close enough that Palmer’s sentence is not hypothetical.

Second, the dividend competes directly with the purchase. Butterfill’s detail from Strategy’s latest weekly update is the one to hold onto: $193 million spent repurchasing STRC, alongside purchases of 267 bitcoin. At Saturday’s price those 267 coins are worth about $21.4 million. The company spent roughly nine dollars defending the funding vehicle for every one dollar it spent on the asset. He adds the sentence that ends the argument: Strategy has already sold some bitcoin to fund preferred dividends.

Sit with that sequence. The vehicle was sold as a way to accumulate bitcoin without selling it. It has now reached the point where the vehicle’s own carrying cost is paid, in part, by selling the thing the vehicle exists to hold. That is not a scandal and it is not insolvency. There is a cash buffer, and repurchases at a discount reduce future dividends, which is defensible capital allocation. It is something more ordinary and more consequential: the machine has stopped adding, and started maintaining.

To see what changed, put the ETF channel beside it. US spot bitcoin ETFs have taken in roughly $57 billion cumulatively and hold around $106 billion in assets, per Butterfill. An ETF share is a claim that extinguishes itself the moment the holder sells. A perpetual preferred is a claim that persists whether or not anyone wants it, and it must be fed in cash. Both are demand. Only one of them can become supply at a moment’s notice, and only one of them stops costing money when the buyer walks away.

One illustration, clearly labelled as an illustration and not as accounting: if the entire $12 billion Fidelity sizes the preferred market at had been raised and spent on those 135,979 coins, the average entry would be about $88,300, roughly ten per cent above Saturday’s price. It was not: the $12 billion is the size of the market today, not the sum ever raised, and not all of it went into bitcoin. But the two numbers sit in the same frame now, and a year ago they did not. The funding was raised against an asset that was rising; the asset is no longer rising faster than the funding costs.

Twenty ounces

There is a second market in the same week that makes the point without any of the leverage.

Gold for December delivery last traded at $4,007.90 an ounce. One bitcoin buys 20.04 ounces of it, my own division of two prices read at 09:30 Central European Time on Sunday.

Brent for December delivery last traded at $111.70 and West Texas at $103.27, both up more than three per cent on Friday, with American personnel being evacuated from embassies and bases across the Middle East on intelligence of Iranian attack planning, as reported by The Wall Street Journal and summarised by CoinDesk. Nik Bhatia of The Bitcoin Layer points at the buffer that would normally absorb a shock like this: the US Strategic Petroleum Reserve is at its lowest level in more than forty years, drawn down through the Iran conflict, and cannot be tapped twice.

This is the part of the argument that bitcoiners should find uncomfortable rather than reassuring. The monetary debasement trade is not dead. It is being paid, in size, to gold and to oil, in a week when the dollar index sat at 101.75 and the S&P closed at 7,623. And bitcoin, the instrument built for exactly this scenario, is 35 per cent off its high and 9.5 per cent down on the year.

I will not pretend to know why, and I mistrust anyone who is certain. What can be said without guessing is that a hedge which falls during the shock it was bought for is, for the duration of that shock, a leveraged position with a story attached. The story may still be right. The leverage is what the weekend measured.

One month

Now the part that will decide whether the AI capex trade can be repaid, and it arrived in the same seventy-two hours.

Moonshot released Kimi K3.1 Pro: a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active, a one-million-token context, and a post-training run the company describes as substantially larger than its base model’s — the whole stack optimised, in its words, so that “each increment of RL compute delivers stable, predictable returns.”

On Artificial Analysis’ Intelligence Index it placed fourth at 55. Opus 5.5 scored 58. Sonnet 5.5 and Muse Wonder Full scored 56. On Arena’s Code Arena for web development it landed eighth at 1,692 points, up nine places and 37 points on Kimi K3. On Artificial Analysis’ Cyber Index it scored 64 per cent, the highest result for a publicly accessible model, second overall behind a trusted-access model, GPT-6 Sol Daybreak Blue, at 69.

Read those three results together and the shape is clear. Fourth overall, eighth in coding, second in offensive cyber capability. That is a frontier-adjacent model. It is also, as of this writing, not open: the API shipped, the weights did not. Moonshot promises them on 4 November, and until that day the honest label is “soon-to-be open-weight”, not “open source”. The license, the redistribution terms, the commercial restrictions, the checkpoint precision and the full serving requirements are all unknown, and Artificial Analysis recorded the weights as unpublished.

Kimi K3.1 Pro against the frontier on the Intelligence Index and the Cyber Index

There is an arithmetic here that kills the local-model fantasy, and it takes one line. Two point eight trillion parameters is about 5.6 terabytes at sixteen bits, or 1.4 terabytes at four bits, before quantisation metadata and serving overhead. That figure is illustrative, a round-trip from parameter count to bytes rather than a measured deployment. The order of magnitude, though, is not in dispute. Self-hostable and consumer-device-local are different properties, and a trillion-parameter model is only ever the first.

Sebastian Raschka’s framing of the same stack, published the day before, is the one I would keep if I had to keep one. The model is an engine; the agent harness is the machinery built around it, and its job is “sophisticated context management and giving the LLM access to tools”. A reasoning model is an upgrade to the engine. The harness is where a company can still be better than its competitor’s engine, and it is also the layer that a one-month model gap does not erase.

Now the economic half, which is the part almost nobody connected to the capability half this week.

DeepSeek published its internal inference stack, DeepInfer, in the same window. Interconnects’ Jack Clark and Nathan Lambert call it one of the bigger open-source artifacts of recent months and note two things about it: it is likely far more efficient than the open serving frameworks everyone else uses, and estimates put DeepSeek’s gross margin on token sales at up to 90 per cent when running on it. It is the first time a leading lab has handed out the tools that produce its margin.

Put the two facts side by side. The capability lead over open weights is now measured in about a month. The cost advantage that the closed labs hold over open serving is now being published as software.

Neither fact proves that hyperscaler capex is wasted. But both narrow the set of places the return can come from. If the model is a month behind, the return cannot be the model. If the inference stack is public, the return cannot be a proprietary cost curve. What is left is distribution, power, physical plant and the switching costs of enterprises already integrated. That is a real moat, and a much less exciting one than the one the equity prices were built on.

Meta opened the week by announcing more aggressive bets on AI investment. Alphabet, Amazon and Microsoft report next week. Into that, the market delivered a near three-sigma move in US rates, a Nasdaq that sold off alongside it, and, per Bloomberg as cited by Bhatia, a call-to-put ratio on long-dated Treasuries that turned positive for the first time since the first quarter. Traders buying upside on long bonds are positioning for a growth scare rather than an inflation one, which is a statement about positioning and not about what happens next.

The bill in three currencies

It is tempting to file these as three stories: a crypto flush, a commodity squeeze, an AI benchmark. They are one story told in three units.

Every structure that expanded this cycle borrowed against appreciation. The treasury companies borrowed against bitcoin’s price, and the loan pays 12 per cent in cash while the collateral pays nothing. The hyperscalers are borrowing against a capability lead, and the lead is a month wide. The leveraged trader borrowed against volatility staying low, and two thirds of a per cent in fifteen minutes was enough to take four hundred million from him.

None of the three is a fraud, and that distinction is the whole of the analysis. Each is a rational structure that works while its margin holds and fails without drama when it does not. The failure mode is quieter than collapse: a slow change of direction, the buyback in place of the purchase, the published inference stack instead of the proprietary margin, the margin call instead of the position.

The same week supplied the code-level version of the same lesson, and it took seven minutes and twenty-one seconds. On Friday an attacker minted two million AUSD on Initia’s layer 1 through what co-founder Zon called a “Day 0 bug” in its MoveVM-on-CosmosSDK integration, burned the unbacked tokens, and pushed them across LayerZero to Ethereum, where the lockbox adapters holding the real assets paid out. By D2 Finance’s breakdown, 1.4 million AUSD left, or 71 per cent of that adapter’s holdings, along with 105 ETH, or 96 per cent of the native-ETH adapter, plus $1.08 million of USDC minted through Circle’s CCTP. Roughly $2.65 million reached Ethereum before the chain was halted. As D2 put it, “an OFT adapter cannot check whether burned tokens were ever backed.” That is a margin too, and it was the assumption that the other side of the bridge existed.

There is a fourth structure in the same week that makes the political point, and it belongs here because it is the one that gets built while the others are being marked down.

Russia’s central bank governor, Elvira Nabiullina, told a 23 October press conference that 270,000 digital ruble accounts had been opened since the nationwide launch on 1 September. Fifteen thousand of those belong to legal entities and sole traders. Roughly 87,000 had opened in the first ten days. She gave no figure for transaction volume, active users or value transferred, and the Bank of Russia still has not published any.

An account count under a compliance mandate measures the mandate. Russia’s twelve systemically important banks were required to offer digital ruble accounts from 1 September, and merchants above 120 million roubles of annual revenue were required to accept it. What is missing from the release is precisely the number that would tell you whether anyone wants the product. That absence is the design, not an oversight: every digital ruble transaction settles on the central bank’s own ledger, the balance earns no interest, and the platform runs smart contracts, which is to say it has both the legal and technical capacity to condition individual units of money. Russia banned crypto for domestic payments while building this. The sequence is the strategy.

The European Central Bank is building the same architecture and describing it differently, which is the honest comparison even where it flatters neither side. In an interview with Milano Finanza published on 24 October, Executive Board member Piero Cipollone set out the timeline: a pilot with real transactions by ECB and national central bank staff in 2027, a formal issuance decision once the regulation is adopted, hoped for by the end of this year, and rollout as early as 2029. Thirty-six market operators, seven of them Italian, are already in the preparatory work. He gave a reason that is not a monetary one: Italian online payments went from 6 per cent of the total in 2019 to 24 per cent in 2024, and two thirds of card payments run over non-European schemes.

His answer on limits is the detail to keep: there will be a cap on how much digital euro a wallet can hold, and no cap on how much can be spent. Whatever one thinks of the project, that is a published design with a number in it, and the number is checkable. Russia published an account count and withheld the only figure that would test it.

The difference between the two rails is not the technology. Both are programmable, both settle on a central ledger, both displace a permissionless alternative by making themselves the default. The difference is whether a citizen, a journalist or a foreign central bank can verify the claim. That is the scarce good in this cycle, and it is the thing the private structures above have been quietly running out of.

The other side of the trade

An argument that only points one way is not an argument, so here is the case against everything above.

Bitcoin at $80,000 is still 37 per cent above its June low, and June was the panic. The preferreds trade at par. The ETFs hold $106 billion in assets. A one-month capability gap is still a gap: the frontier moved repeatedly in the past month, and the compute required to stay in front is not something an open release can conjure. DeepSeek publishing its inference stack lowers the price of tokens, which increases token consumption, which increases demand for compute. That is the Jevons move which has rescued every cost-deflation story in this industry so far. And the treasury companies are not passive: a 12 per cent coupon on a perpetual is a claim on a company that can buy its own paper at a discount with money it raised at par, which is accretive in its own right.

The strongest version of the bear case against me is this: none of the three margins has actually broken, and markets that price a break that does not arrive tend to punish the people who called it. That is true, and it is why I have not written a single forward price in this piece.

What I will defend is narrower. The direction of each structure has changed, measurably, in the last ten days. The purchase-versus-buyback ratio flipped. The ETF flow turned negative after three days of inflows. The open-weight gap printed at one month, on an independent index, with the weights not yet out. And the debasement bid that was supposed to arrive at bitcoin arrived at gold instead, at $4,008, while bitcoin lost 9.5 per cent of the year.

What would change the answer

Watch the ratio, not the headline.

For the treasury trade: whether STRC holds near $99.40 through the next weekly update, and whether the bitcoin line in that update is larger or smaller than the repurchase line. If the company is still buying paper instead of coins two updates from now, the machine is in maintenance and the market will eventually price the dividend as a cost rather than a feature.

For the AI trade: whether the one-month gap holds at the next release. One month is a data point. Two consecutive months is a structure, because it would mean the open side is tracking the frontier rather than chasing it, and the capex has to be repaid out of something other than capability.

For the leverage: open interest at $140.10 billion against a weekend of $815 million in forced selling. That is a structure sized for a market that moves less than this one does.

For the rails: whether the digital euro regulation is adopted this year, and whether the Bank of Russia’s next release contains a transaction figure alongside its account count. One of those two institutions will tell you what it is doing. That is the tell.

Bitcoin fell through $80,000 on a Saturday because a levered book could not absorb two thirds of a per cent. That is not the story. The story is that the two structures which bought the most bitcoin and the most compute over the past two years were both financed on margins that were thinner than the pitch deck said, twelve per cent in cash and one month in capability, and this was the week the market started asking what they cost to keep.

What I could not check

I did not read Strategy’s or Strive’s filings. The $193 million buyback figure and the 267 coins are CoinShares’ characterisation of a weekly corporate update, relayed by CoinDesk, and I am one step removed from the primary document.

The liquidation data is CoinGlass, as reported by The Block. I did not query any exchange’s feed directly. The twelve-month price levels are my own pull, and a different venue’s daily close would move the percentages by a few tenths of a point.

The 5.6 and 1.4 terabyte figures are a conversion from parameter count at sixteen and four bits, with no allowance for what serving actually costs. Nobody has measured a K3.1 Pro deployment in public, because the weights are not out.

I did not obtain Artificial Analysis’ methodology, only its published scores. And there is no digital ruble transaction data to check, in either direction, in any source.

Sources

Write a comment