Fifteen Percent for a Promise

Three numbers on one screen, same currency, same morning.

The ten-year Treasury pays 5.069 percent. A bond issued by CoreWeave, the company that rents out graphics processors to artificial intelligence developers, has been changing hands at fifteen. And an ordinary depositor at a large American bank is paid something in the neighbourhood of 1.4 percent for leaving cash where the bank can lend it out.

Those look like three different investments. They are really three prices for the same object: a promise about the future. The promise is not equally good, and this year the market has started saying so out loud. What follows is an attempt to put the ladder in order — from the promise that pays the least to the one that pays the most — and then to answer the question that kept interrupting me while I was doing it. Why is the one asset on the screen that is not a promise at all the one that fell?

Bitcoin traded at $82,789 when I checked, up on the day, after dipping below $80,000 over the weekend. James Check, who has spent years building on-chain frameworks for exactly this kind of week, opened his Monday note with the honest observation that sentiment now swings with the colour of each daily candle. He is right, and that is the tell. If the week's story were about Bitcoin, sentiment would not be the interesting variable. It is about the price of every other promise in the stack, and Bitcoin is where the leverage gets liquidated first.

The ladder, in order#

Start at the bottom, with the cheapest promise money can buy: a bank deposit. Mel Mattison, a former private-wealth manager, walked through the mechanics on TFTC this week and the arithmetic is not complicated. Bank of America's own reported cost of deposits runs around 140 basis points. The bank takes the money, lends it or parks it at the policy rate, and keeps the difference. That spread is not a fee anyone agreed to. It is the price of a settlement system that takes days and therefore forces people to hold a buffer at zero or near-zero interest.

Mattison's defence of the practice is not that it is fair. It is that it is invisible, and that the invisibility is purchased. The banking lobby, in his telling, stripped the interest provision from stablecoin legislation and killed the market-structure bill outright, because a dollar token paying something near the bill rate makes a checking account pointless. Lummis's three-part plan scores, by his count, roughly 0.75 out of three. Whatever one thinks of the politics, the economic logic is airtight: interest-bearing digital dollars at scale are an existential product problem for deposit franchises. Nobody lobbies that hard against something harmless.

Then comes the Treasury. I measured 5.069 percent on the American ten-year this morning. The gilt sat at 5.374, the BTP at 4.524, the Bund at 3.412, the JGB at 2.904 — six sovereigns, one direction. That is a global repricing, not an American one, and it has been going on long enough that the word "crisis" has worn out.

Above the Treasury sits high-yield credit, roughly eight percent for the broad American junk market. And above that, at the top of the ladder, sits the paper of a single company in a single industry, at fifteen percent, unsecured and rated junk.

Nine hundred and ninety-three basis points. That is the distance between lending to the United States government for a decade and lending to the company building the physical layer of the artificial intelligence boom. Three times the sovereign yield, for the privilege of being earlier in the queue for a cash flow that does not yet exist.

The arithmetic CoreWeave cannot argue with#

Cory Klippsten laid out the numbers from the company's own reporting, and they are stark in a way that spreads are not. In the first half of this year CoreWeave spent fourteen billion dollars on equipment and infrastructure. It generated less than four billion in operating cash. Call the gap ten billion; it is about seventy-one percent of the spend, and it was covered mostly by borrowing.

This is not fraud or folly. It is the shape of the business. You build the capacity your contracts call for, you spend first, and the revenue arrives over the following years. Every capital-intensive industry has done this — railways, fibre, fabs. The question is never whether the spending precedes the revenue. It is whether the revenue arrives before the next lender asks for more collateral.

And that question is being answered in the credit market right now, not in the equity market. A fifteen percent yield on unsecured paper does not mean CoreWeave pays fifteen percent on all of its debt; existing fixed-rate coupons do not reprice. It means the marginal lender has repriced, and it means the next loan — the one that finances the next data centre — arrives with a higher rate, more collateral, and more of the company's own equity required. Klippsten's point is that this is how a build-out slows without a single bankruptcy: the project that no longer clears its cost of capital gets delayed, and a delayed project is somebody else's cancelled order. The chipmaker's forecast, the contractor's backlog, the switchgear factory's delivery schedule.

The macro version of the same arithmetic comes from PIMCO's September consensus figures, which put next year's combined hyperscaler cash shortfall after capital spending at eighty-seven billion dollars. The largest companies in the world generate extraordinary cash and are still outspending it. They can fund the gap. The question is at what price, and what that price does to everyone standing behind them.

Revenue per gigawatt is the only number that decides this#

Here is where the analysis usually stops, with a debt yield and a shrug. The better version keeps going, because there is exactly one ratio that determines whether the fifteen percent is a bargain or a warning.

Revenue per gigawatt. How much money does an AI lab make from each unit of data-centre power it consumes?

The Pomp Letter's chart set this week put OpenAI at roughly ten billion dollars per gigawatt, a figure that has held steady since 2023. Anthropic's run rate passed sixty-five billion in July without disclosing capacity; if it runs two gigawatts the figure is about thirteen billion per gigawatt, if five gigawatts then thirty-two and a half. Same order of magnitude, both of them, and that is the good news.

Now put the cost side next to it. Since the end of 2021 the price of switchgear — the equipment that routes power inside a data centre — is up about sixty percent. Transformers, cooling equipment and generators are each up roughly thirty. General construction materials about ten. Memory for data centres is more than five times the price of a year ago, which is why Apple said in June, with unusual directness, that it had never seen a component price increase this large this fast. Power equipment makers have backlogs stretching past 2030.

So the numerator is flat and the denominator is inflating. That is the entire story of the fifteen percent, and no amount of demand enthusiasm changes it. Demand does not pay for capacity; price does. The build-out has been financed on the assumption that revenue per gigawatt would rise as models got better and customers got locked in. Instead the models got cheaper — open-weight models now handle seventy-eight percent of the tokens on one major AI gateway, by the Pomp Letter's count, while closed models still take more than half the dollars because they are reserved for the hardest work. Cheap tokens at the bottom, expensive tokens at the top, and a capital structure built for volume at the top.

Which brings up the detail I keep returning to. Only 492 billion dollars of the 2.3 trillion-dollar data-centre pipeline is under construction. The remaining 1.8 trillion will be built at today's inflated prices, financed at today's credit spreads. That is not a forecast so much as a queue with a delivery date nobody has fixed.

The honest counterargument#

I do not want to write the bear case and call it analysis, because the strongest evidence in this week's material points the other way, and it comes from someone who has been consistently right about the bond market this year.

Nik Bhatia's argument, made across two notes, is that the Treasury market is not bidless and never was. Look at the auctions, not the yield. At the last reopenings before November's quarterly refunding, indirect and direct takedown was strong, particularly in tens. Foreign and international participants have increased their share. And primary dealers — the market makers who absorb whatever is left over — are at record-low allocations. That last number matters because it inverts the usual reading: a low dealer share means the dealers were not forced to eat the issue. The buyers showed up.

His explanation is mechanical rather than narrative. At five percent and above, a Treasury offers three things simultaneously: positive carry against repo financing for leveraged buyers, a five percent coupon as downside protection, and positive convexity as you move down a yield curve that starts paying you more the more rates fall. Price-sensitive holders do not need to be brave at 5.3 percent. They need to be patient, and patience is what the bid since early October looks like from the outside, even as yields rose.

And then the part that sounds wrong until you sit with it. In the TBL framework, a stable and strengthening Treasury market expands system liquidity, because the collateral at the base of the credit system becomes easier to rehypothecate. A strong bond market is not a flight to safety; it is an increase in the system's capacity to lend against itself. That is why his liquidity index is turning the way it did in early April, which was followed by an upturn in both bitcoin and equities.

I take this seriously because it is falsifiable and it has been right. The bear case I have been building above is not about whether the Treasury can fund itself. It can. My case is about the layer above: whether the collateral chain built on top of that base — the private credit, the convertibles, the take-or-pay contracts, the fifteen percent paper — is priced for a world where the base stays stable. CoreWeave's debt trading below eighty cents on the dollar says the market has begun to doubt the layer, not the base. Those are different claims, and conflating them is how people end up confidently wrong in both directions.

The sovereignty multiple#

Here is where I want to put a number on something that usually stays a feeling.

Cory Klippsten asked his readers a question that is almost rude in its directness: would you trade two million dollars of a Bitcoin ETF for one million dollars of bitcoin you hold yourself?

Whatever ratio makes you indifferent, that is your sovereignty multiple. His own is about 1.5, he says, and he is an American with access to the American financial system and good lawyers. He has said that in a country with serious financial repression his multiple would be ten. He has friends, in places where bitcoin is the only asset they trust, whose multiple he describes as effectively infinite. When he polled the question publicly, more than half of respondents put themselves above five.

I find this the most useful idea I read all week, because it converts a philosophical argument into a discount rate. Two million of paper for one million of possession means you price control at a fifty percent haircut. Above five means you would surrender eighty percent of the nominal value to keep the thing itself. That is not a market price. It is a personal one, and it varies with how much you trust the institutions holding your money — which is precisely why it cannot be arbitraged away and why it is not a bubble signal either.

Klippsten's sharpest point is that a solvent bank can still say no. Your account can show a million dollars while you are somewhere else wondering how to pay the rent. Nothing has disappeared. The claim simply stopped being performative where you needed it to perform. That is not a credit event; it is a jurisdictional one, and no credit rating captures it.

Why the thing with no counterparty falls hardest#

Which returns me to the question I started with. Every instrument on the ladder is a claim on somebody. The deposit is a claim on a bank. The Treasury is a claim on a tax base. The junk bond is a claim on a company's future cash flow. Bitcoin is not a claim on anything. It is the only line on the screen where the counterparty is nobody, and this month it is the line that fell.

My reading — and I want to label it as a reading, because the data does not settle it — is that this is what a liquidity preference looks like from the inside. When credit spreads widen and the marginal lender reprices, the first thing institutions sell is not what they believe least in. It is what they can sell fastest, at the least cost, without negotiating with a counterparty. That is the asset with no counterparty. The same property that makes bitcoin a hedge in a monetary crisis makes it an exit in a margin crisis. It is not contradiction; it is the same attribute used in two different weathers.

There is an irony here that the ETF wrapper made sharper. The wrapper that gave millions of people access also gave them a button. One click and the exposure is gone — and with it, the thing that made the exposure worth holding. The sovereignty multiple is, in a sense, the price of not having that button. Klippsten's two-for-one trade is someone deciding that the illiquidity is the feature.

For the record, on Bhatia's own numbers: his indicator confirmed red on September 11 at $76,600 for bitcoin and $761.53 for the S&P 500 tracker. Bitcoin at $82,789 this morning is about 8.1 percent above that level — my arithmetic on his stated entry, not his own figure, and he notes the non-trade missed as much as thirteen percent at moments. Equities are up roughly twelve percent on the year. The signal has not been a disaster. It has been a reminder that being out of a market is also a position, and it has a cost.

The physical price of possession#

If possession is worth a premium, it also has a bill, and this week someone finally sent it.

Jameson Lopp's archive of physical bitcoin attacks crossed seven hundred documented incidents on 16 October. A researcher working as Pizza Bitcoiner used agentic AI searches to sweep local-language news, court records and police reports across seventy-two countries, and surfaced 375 cases that had never appeared in the English-language record. The archive roughly doubled. Lopp's own summary was characteristically dry: the good news is the archive got more accurate, the bad news is we just doubled the known attacks.

I did not want to take the total on faith, so I pulled the repository and counted the rows myself this morning. Seven hundred and ten entries. The year-by-year shape is the interesting part, and it is not the shape most people assume.

Four hundred and fifty-five of those incidents are dated 2023 or later — 66 in 2023, 99 in 2024, 165 last year and 125 so far this one, all counted from the repository myself. France accounts for eighty-five, more than any other country, ahead of the United States at seventy-seven and Taiwan at fifty-six. That geographic order is not what an English-language reader would guess, which is exactly the point: the archive's previous count of roughly 350 was not a measurement of the world. It was a measurement of one researcher's reach into it.

Here is the contradiction that makes the epistemics concrete. Press accounts published in mid-October, describing this very archive, put last year at "seventy-plus" incidents and called 2025 a record. Counting the repository myself gives 165 for 2025. Same archive, same year, two numbers differing by more than a factor of two. The rows do not record when they were added, so I cannot prove from the file alone that the difference is entirely retroactive backfill — but nothing else explains a doubling of a completed year.

That is the lesson, and it applies to every number in this essay. The frequency of attacks did not double in October. The lens widened. A dataset built on English-language search was reporting the English-speaking world and calling it the world, and the gap between those two things was 375 violent incidents. Anyone who built a threat model on the old number was calibrated to a sample that was at least half empty.

The threat itself is not in doubt, and it is not evenly spread. It concentrates on people whose on-chain holdings are traceable to a face: conference speakers, executives who disclose, posters who name their wallets. Lopp's standing advice is structural rather than behavioural — arrange custody so that immediate compliance under duress is physically impossible. Time locks, geographically distributed keys, decoys. Not a better hiding place. An inability to hand it over.

What is actually being repriced#

Put the pieces together and a single sentence comes out.

The world is repricing the cost of believing somebody else. A depositor is paid 1.4 percent to believe a bank. A Treasury holder is paid 5.069 percent to believe a government. A junk buyer is paid eight percent to believe a balance sheet, and a CoreWeave lender is paid fifteen to believe a business model whose numerator is flat while its denominator inflates. Each of those numbers is a measurement of trust, taken daily, in public, by people with money at stake.

And then there is the one asset that asks you to believe nobody, priced this morning at $82,789.

The interesting thing about that ladder is not that the top rung is expensive. It is that the rungs are diverging. A world in which sovereigns, corporates and deposits all paid roughly the same was a world that assumed the promises were interchangeable. That assumption is gone, and it is not coming back on its own. What replaces it is a hierarchy, and hierarchies of credit are uncomfortable places to run a leveraged build-out.

The honest part: what I could not verify#

Three things I cannot demonstrate, and will not pretend to.

First, there is no evidence yet of the breakdown I have been describing. CoreWeave raised more than four billion dollars in convertible bonds in September. The hyperscalers have cash flows and options that most companies do not. A fifteen percent yield on one unsecured tranche is a signal about the marginal lender, not a verdict on the borrower. If financing costs stabilise here, the build-out continues and the fifteen percent looks like a buying opportunity in hindsight.

Second, bitcoin's fall may be nothing more than beta. It is a high-volatility asset that trades around the clock and gets sold first when risk budgets shrink. I have offered a reading of why the selling concentrates there; I have not measured it, and a simpler explanation fits the same facts just as well.

Third, my attack counts are documentation, not incidence. Seven hundred and ten rows in a public list is a floor built by volunteers and one AI pipeline, not a census. The archive's own falsifiability test — an independent audit finding more than about twenty percent duplication among the 375 new cases — has not been run. Until it is, the doubling is a better-sourced estimate, not a settled fact.

What I will stand behind is the arithmetic. Fourteen billion spent against less than four billion generated. Ten billion dollars of revenue per gigawatt that has not moved since 2023, against switchgear up sixty percent and memory at more than five times its price a year ago. One point eight trillion dollars of pipeline still to be financed at those prices. And a credit market that has just told one participant, in the only language it speaks, what that combination is worth.

The promise that pays the least is the one you can spend tomorrow without asking anyone. The promise that pays the most is the one nobody can verify. Everything between them is a negotiation about time, and this year the negotiation got more expensive.

Bitcoin's contribution to that argument is not that it went up. It is that it is the only thing on the ladder with nobody on the other side of the trade — which is why the sovereignty multiple is not a slogan, but a number each holder has to arrive at alone, in whatever currency their own jurisdiction happens to speak.

Sources#

• Checkonchain — James Check, "Assessing Pullbacks", 19 October 2026
• The Daily Bitcoiner — Cory Klippsten, "What Happens to the AI Build-Out When the Next Loan Gets Harder to Obtain?", 16 October 2026
• The Daily Bitcoiner — Cory Klippsten, "Would You Trade $2M of a Bitcoin ETF for $1M of Real Bitcoin?", 17 October 2026
• The Bitcoin Layer — Nik Bhatia, "TBL Liquidity Indicator Signal Update", 17 October 2026
• The Bitcoin Layer — Nik Bhatia, "What If US Treasury Yields Go Back To 3%?", 15 October 2026
• The Pomp Letter — Nick Fouriezos, "Five Friday Charts: Where AI money really goes", 16 October 2026
• The Pomp Letter — Nick Fouriezos, "Are supply chain oligopolies the better AGI bet?", 15 October 2026
• TFTC — "Mel Mattison: Banks Are Fleecing You", 17 October 2026
• TFTC — "Lopp's Wrench Attack Archive Doubles to 702 After AI-Assisted Search", 17 October 2026
• GitHub — Jameson Lopp, Known Physical Bitcoin Attacks (repository counted directly, 710 rows, 19 October 2026)
• Pizza Bitcoiner — Physical Bitcoin Attacks database
• CNBC — US 10 Year Treasury yield, read 19 October 2026, 09:30 CEST