The Debt Nobody Can See

Two central banks admitted on the same day that the AI buildout runs on debt they cannot measure, and both kept policy pointed at higher rates. That is not a dilemma, it is a maturity mismatch, and it has a date.
The Debt Nobody Can See

The Debt Nobody Can See

On Wednesday 18 November, at 2 pm Washington time, the Federal Reserve published the minutes of its October meeting. A few hours earlier the European Central Bank had published its Financial Stability Review. Neither institution coordinated with the other. Both wrote the same paragraph.

The Fed, on the AI buildout: “Most participants observed that AI-related debt had grown rapidly and that much of this growth had taken place through complex and opaque financing structures.” Fed staff added that “off-balance-sheet leasing commitments and circular financing arrangements among AI-related firms could amplify losses if returns on AI investment were to disappoint.” The ECB, in its own words: “Investment in AI infrastructure is increasingly funded through debt, such that a potential change in sentiment could spread losses from equity markets to bond markets.”

These are the two institutions that set the price of money in the two largest currency areas. Within one day they both put on paper that the largest capital expansion of the decade runs on borrowed money whose structure they cannot fully measure, and both said the right response is to keep the cost of that money rising.

That combination is usually reported as prudence. A maturity mismatch with a date attached describes it better.

What the committee actually wrote

The headline was a hold, and the headline was wrong. The target range stayed at 3.75 to 4.00 percent on a 10 to 2 vote, and the minutes show that the vote understated the direction of travel. “Some participants favored an increase of 25 basis points in the target range at this meeting, emphasizing that inflation had remained too high and that they saw few signs of it abating.” A few of those who voted to hold called it “a close call.”

Then the sentence that matters for December: “almost all participants judged that, under their baseline outlooks for inflation and the labor market, further action to raise the target range for the federal funds rate … would likely be appropriate.” A couple went further and suggested that “an increase of more than 25 basis points could warrant consideration at the next meeting if incoming inflation data were unfavorable.”

The next meeting is 8-9 December, with a Summary of Economic Projections attached, according to the Fed’s own calendar. Some of the commentary circulating that morning put it at 9-10 December and put the minutes release a day late. The primary document is three clicks away and takes minutes to read. That is the standard this column holds itself to, and it is worth applying to the loudest voices too.

So: a pause that was pre-announced as a pause, with tightening priced as the next move, adopted while the same committee described a credit structure it cannot see.

The half that is not in the statement

Central bank risk paragraphs are usually boilerplate. This one is not, and the reason is the last sentence of the passage that the wires did not quote.

“Several participants expressed concerns that the existing data available to bank supervisors may not fully capture banks’ indirect exposure through nonbank financial institutions, including credit lines to private credit funds and other complex linkages.”

Read that again as a supervisor. The exposure sits at a bank. The loan is not to the AI company. The loan is to a fund, which holds a note issued by a leasing vehicle, which owns the servers that a model company uses under a contract that may or may not be a capital lease. The bank’s quarterly return sees a line of credit to a fund. The risk sees a data centre with a fifteen-year life and a customer with no fifteen-year history.

The ECB named the same mechanism from the other side. Its vice-president, Boris Vujčić, put the asymmetry precisely: “Banks are well placed to withstand shocks, while vulnerabilities are more pronounced in certain non-banks where risks are growing and macroprudential safeguards are less developed.” Its review adds the detail that matters for a bond portfolio rather than an equity portfolio: valuations “rely on ambitious earnings expectations which could yet be challenged by financing concerns or by competition from cheaper AI models.”

Two central banks, two days of drafting, one conclusion. The leverage is outside the perimeter, and the trigger is not a fall in the share price. It is a refinancing.

Why 120 basis points is the whole story

Here is the arithmetic nobody in either document writes down, measured on the tape this morning.

The ten-year Treasury yield was 5.076 percent at 09:31 CET on 19 November, against a policy range whose midpoint is 3.875 percent. That is roughly 120 basis points of term premium over the policy rate, my own calculation on my own readings (19 November 2026, 09:31 CET). The long end is already pricing the fiscal half of the picture the ECB described, and it is doing so before the Fed has hiked again.

Now put a data centre against that curve. The asset has a fifteen-year useful life, a five-year customer contract, and, in the current fashion, a financing structure that reprices every few years. The revenue is denominated in tokens. The cost is denominated in basis points. Every quarter that the policy rate rises moves the cost of the liability and leaves the revenue unchanged.

That is the mechanical difference between an equity bubble and a credit bubble. An equity bubble deflates: the price falls, the disappointment is marked, a new cohort of buyers arrives at a lower price, and the buildings stay standing. A credit bubble cannot deflate, because the claim is fixed in nominal terms. It either gets refinanced or it gets liquidated. The Fed’s own staff used the word that gives the game away: “amplify losses.”

The case study is not hypothetical

On the same day as the minutes, a company called Hyperscale Data filed a disclosure that reads like a diagram of the mechanism.

Hyperscale was a bitcoin miner. Since late July it has sold its entire treasury, 1,113 BTC for roughly $74 million, including a final 211 BTC for about $16 million announced last week. The proceeds went into debt reduction and into a data centre in Michigan built for an AI customer. That customer agreement “could generate $1.2 billion”, but only if two five-year extensions are exercised and only if the company meets its obligations, including finishing the building.

At the end of September it held $17 million of unrestricted cash against $231 million of current liabilities. Short-term liabilities exceeded current assets by $128 million. Cash had more than halved from $36.8 million at the end of June. It now says it needs another $74.8 million between November and March to complete the project, and it warned of substantial doubt about its ability to continue as a going concern (CoinDesk, 18 November, citing the filing).

Every number in that paragraph is from a public filing. None of them is a forecast. And the sequence is the sequence: mine a volatile asset, sell the asset at a loss to fund a fixed-cost building, sign a contract whose headline value depends on extensions that have not happened, then discover that the last $75 million is the hardest money you will ever raise.

This is what the Austrian tradition calls malinvestment, and the term is often misunderstood as a moral judgement. It is not. It is a statement about the shape of the capital structure. Capital was sunk into a specific, immovable, long-lived asset on the basis of a price signal that was itself distorted by credit conditions. When the signal reverses, the asset cannot move. The loss is not the difference between two prices. It is the difference between what the building can earn and what the paper demands.

The second blade, which is falling, not rising

The Fed minutes focus on the cost of money. The more interesting risk is on the revenue line, and it was visible in the same week’s product releases.

List price per million output tokens and the Artificial Analysis intelligence index for five frontier models, November 2026

Google shipped Gemini 4 Neon at $3 per million input tokens and $15 per million output tokens, with a one-million-token context window and four effort levels. On Google’s own comparison slide, Neon scores 61 on the Artificial Analysis intelligence index against Carbon’s 60. Carbon is currently discounted to $2 per million input and $10 per million output, a promotional price that doubles to $4 and $20 on 1 December. Hydrogen, the cheap tier, sits at $1.50 and $7.50 with an index score of 54. The release cadence has compressed to 28 days between Argon and Carbon, 15 between Carbon and Hydrogen, and six between Hydrogen and Neon, by Arena’s count as reported by Latent Space on 18 November.

OpenAI’s position is the opposite bet on the same curve. GPT-6.2 Astra lists at $8 per million input and $40 per million output, with Sol at $2 and $10, and Astra scores 62 on the same index against Anthropic’s 65. OpenAI’s own figures, as reported by Latent Space on 17 November, show why that sticker price is misleading: on GDPval v2 Astra used 22 percent fewer output tokens than its predecessor, and on the Agents’ Last Exam it gained 5.2 points over the previous model at less than a third of the cost per task.

Put the two facts together. The industry is selling a product whose price per unit falls with every release, produced by a capital stock that is fixed, financed with debt that reprices, and depreciates on a schedule set by silicon rather than by demand. The ECB wrote exactly this down: earnings expectations “could yet be challenged … by competition from cheaper AI models.”

The honest counter is that lower prices expand the market, and volume has grown fast enough so far to swamp the price decline. That has been true for two years, and the capital markets agree with it: Anthropic priced its initial public offering at $83 a share, above the range it had been marketing, according to Latent Space’s account of 18 November. The steel-man deserves to be stated in its strongest form. If inference demand grows faster than price falls, cheaper tokens are a margin expansion for whoever sells compute, and the debt is comfortably serviced. The problem with that argument is that it requires the demand curve to keep bending upward at exactly the rate the capital structure assumed, in a market where a competitor can halve the price in six weeks. A business that must grow to service its debt is not a growth business. It is a pyramid with a depreciation schedule.

Binance BTCUSDT perpetual open interest in bitcoin and in dollars, 6 to 19 November 2026

What bitcoin’s tape says that the commentary does not

Against that background, the bitcoin tape this week is more informative than the price.

Bitcoin traded at $63,659.84 at 09:31 CET on 19 November, 49.6 percent below the all-time high of $126,210.50, my own calculation. Gold was $3,737.70, which puts one bitcoin at 17.0 ounces. In a week when two central banks described a debt structure they cannot map, the asset that competes with gold on the monetary axis is trading at a seventeenth of the metal.

The leverage, however, has already left. Open interest on Binance’s BTCUSDT perpetual peaked at 101,152 BTC on 13 November and fell to 92,007 BTC on 18 November, a decline of 9.0 percent in five days. In dollar terms the same book went from $7.43 billion on 10 November to $5.97 billion on 18 November, down 19.6 percent. It has since rebuilt to 96,298 BTC. All of these are my own readings of the exchange’s public open-interest history on 19 November 2026.

That is what a flush looks like from the inside: not a change of mind among holders, but the removal of the people who were renting the position. And the funding rate this morning was 0.006067 percent per eight hours, which annualises to 6.64 percent at three settlements a day, my own arithmetic. Sygnum’s reading, reported by The Block on 19 November, is that funding stayed negative for 67 consecutive days into May before recovering to about 8 percent annualised by the end of October, and it expects spot-and-short-futures carry to beat Treasury bills on average in 2027. The carry is not dead, it is repricing.

Unspent bitcoin by output type at block 975,019, with the share whose public key or script is already on chain

Thirty-five percent

The other number worth having in your head this week is not a price either. It is a census.

At block 975,019, of the 20,109,207.36 BTC sitting in unspent outputs, 7,164,754.12 BTC, or 35.6 percent, sat in outputs whose public key or script was already visible on chain. The author, publishing on Delving Bitcoin, breaks it down. Of the exposed total, 1,937,232.46 BTC sits in P2PK, P2TR and bare multisig outputs, which are public by their form. The remaining 5,227,521.66 BTC sits in P2PKH, P2WPKH, P2SH and P2WSH outputs that count only because the underlying program appears in a table of hashes previously revealed on chain. Of those, 3,301 hashes holding 49,154.61 BTC had no table match at all, so a spend from either address form would put the key on chain for both.

The chain tip at 09:31 CET on 19 November was block 976,557, which puts the census 1,538 blocks, roughly eleven days, behind the head of the chain (own measurement, 976,557 against the census block of 975,019).

What this does not mean is that a third of the supply is about to be stolen. Exposure is not vulnerability, and P2TR reveals a key by design, in the taproot output itself, without exposing anything that a quantum adversary could use today beyond what a thousand other facts already reveal. What it does mean is that the supply is auditable at a resolution that no other asset class can match, and that a meaningful fraction of holders have made an avoidable mistake: address reuse. The number is not a threat. It is a discipline report, published by a stranger, verifiable by anyone with a node.

Now set that against the central bank paragraphs from Wednesday. The Fed said it cannot see the credit. The ECB said the losses could migrate from equity to bonds. And here is a ledger where every satoshi is locatable, whose inflation schedule is a consensus rule, and where the only thing that changes without anyone’s permission is the price. That contrast is the actual content of the “digital gold” argument, and it has nothing to do with the price chart.

The same problem, in the other room

There is a second document from 18 November that belongs in the same drawer, because it is the same failure of visibility.

The US Government Accountability Office published a covert audit, GAO-27-108814, requested by Senator Maggie Hassan. Investigators posing as ordinary users bypassed the safeguards of all four major chatbots using common techniques and obtained detailed, expert-verified attack plans, for a mass-casualty device and for an assault on grid infrastructure, in under twelve hours each. The GAO’s own framing is the important part: “The details behind the information provided by the AI platforms were not novel. However, the generation of plans that could have taken weeks or months were instead completed in hours with the assistance of AI.”

No new knowledge was created. The time and skill required to act on existing knowledge collapsed. All four companies acknowledged the finding and argued that the tests used spring 2026 model versions. That may be true. It cannot be checked from outside, because every one of these systems is a policy layer inside a proprietary black box that one company updates on its own schedule. “Trust us, we patched it” is a press release, not a security architecture, and it is the same sentence the Fed and the ECB declined to write about the credit stack.

The reflex in Washington is pre-release certification. That would be a licensing regime for thought, drawn up by the four firms that just failed the test, with compliance costs that only they can absorb. The honest tension is that open weights face the identical jailbreak with no vendor to patch at all. Neither openness nor licensing solves the underlying problem, which is that a guardrail implemented as a filter in front of a model is not a property of the model. Verification is the scarce good, in AI and in credit alike.

The honest part

Bitcoin fell 49.6 percent from its high while the credit story was developing. It did not decouple, and anyone who told you it would has some explaining to do. It traded as a liquidity asset, sold by the same desks that sold everything else when the margin call arrived, and the observation that its one-year volatility floor has stayed under 47 percent in this drawdown, against 65 percent in 2022 and 86 percent in 2018 by Strive’s reading as quoted by Pomp on 18 November, cuts both ways. Calmer downside means a broader holder base. It also means bitcoin is becoming more correlated with the rest of a portfolio, not less, and the day it stops being a diversifier is the day it starts being a position.

The Austrian case for bitcoin does not rest on it rising during a credit event. It rests on the fact that it has no refinancing date, no counterparty who can call it, and no board that can dilute it. In a week when two central banks admitted they cannot map the debt in the largest capital boom in a generation, that is not a small property. It is the one property on the list that does not require anyone’s permission.

What to watch, and when

Four dates carry most of the information.

On 1 December, Gemini 4 Carbon’s promotional pricing expires and the input price doubles from $2 to $4 per million tokens. Watch whether demand holds at the higher price. That is a live experiment on the elasticity that every AI credit model assumes.

On 8-9 December the FOMC meets with a new Summary of Economic Projections. The minutes have already told you the direction; the projections will tell you the destination, and the December statement will tell you whether Governor Waller’s public call for a half-point move, as reported by TFTC on 18 November, found a second.

Through the winter, core CPI against the 2.5 percent line that TFTC named as the tripwire for its own thesis. Three consecutive core prints at 0.3 percent have already arrived, on TFTC’s count. A fourth turns “higher for longer” into a policy, and a policy into a refinancing schedule for every structure the Fed says it cannot see.

And before the end of March, Hyperscale Data either raises $74.8 million or it does not. That is one company, and it is small. It is also the cleanest test available of whether the capital markets will keep funding the transition from hashing to hosting at these rates. If the answer there is no, the question moves up the stack quickly, because the same lenders are exposed to the same structure through the funds the supervisors cannot see into.

The boom will not end because sentiment turns. Sentiment turns because the boom ends, and the boom ends on a date that was printed on the note when it was issued. That is what a central bank discovers when it holds rates above the return on capital for long enough: not a recession, but an inventory of promises coming due.

Sources

Write a comment