Seigniorage

Section one

The tell

In June, gold passed US Treasuries to become the largest reserve asset in the world.1

Reserve assets are what a country's central bank keeps its national savings in. For eighty years the default answer has been US government bonds: they pay interest, they can be sold instantly in any quantity, and the United States has never failed to pay. Gold pays nothing and costs money to store. It has overtaken them anyway.

Central banks are the least exciting buyers in existence. They are slow, political, and averse to looking foolish, and they hold positions for decades. They are also the best-informed judges of government credit alive, for the simple reason that they are government credit. When that group moves weight out of the safest asset in the world and into a metal, it is not a trade. It is a revised assumption.

For forty years the price of gold was well explained by a single number: the real interest rate.

The real interest rate is what a bond pays after inflation takes its share. A bond paying 5 percent while prices rise 3 percent returns about 2 percent in real terms. That is the number gold competes against, because gold pays nothing at all: the more a bond returns in real terms, the weaker the case for a metal sitting in a vault.

The rule that follows is simple. Real rates up, gold down. It held through the eighties, the nineties and the two thousands. In 2026 the ten-year real rate rose about forty basis points, meaning bonds became more attractive.2 Gold should have fallen. It rose thirty-three percent.3

Run today's real rates through that relationship and it returns a gold price near thirty-five hundred dollars an ounce. Gold trades at forty-five eighty. The gap is roughly a thousand dollars wide, and it is the subject of everything that follows.

The obvious explanation is inflation. People buy gold when they expect money to lose value, so perhaps that is all this is. If so, it would appear in the instruments built specifically to price inflation, and it does not.

The breakeven rate is the bond market's own inflation forecast. It is the gap between an ordinary bond and an inflation-protected one of the same maturity, and it gives the rate of inflation at which the two would pay the same. If traders expected serious inflation, this number would rise.

In 2026 it rose seven basis points.4 Seven, during an energy war, is a rounding error. So the bond market is stating calmly that it does not expect much inflation, while the gold market is stating urgently that it wants out of the currency regardless. Those two positions are compatible in exactly one way: what is being avoided is not a rate of price increases. It is the institution doing the issuing.

Section two

The machinery

On 26 August the United States crossed forty trillion dollars of federal debt.5 The deficit is running near two trillion a year. Interest alone costs about a trillion, which is to say the government now spends more servicing what it borrowed in the past than it spends on almost anything it is borrowing for in the present.

None of that is the interesting part. Large debts are ordinary, and countries carry them for decades without incident. What is new is the response to them.

In August the government held one of its regular auctions of thirty-year bonds, which is how it borrows for long periods. To attract buyers it had to promise 5.216 percent a year, the highest rate it has had to offer in twenty five years.6

When borrowing becomes expensive, the conventional response is to borrow less of it. The Treasury did three other things instead.

It held the size of its long-term auctions unchanged rather than shrinking them. It shifted its borrowing toward short-term IOUs, which cost less today but must be repaid or rolled over within months, so the problem is deferred rather than reduced. And it doubled its purchases of its own outstanding bonds in the open market, which raises their price and lowers their yield.

The third of those is the significant one. The government entered the bond market and bought its own debt so that the price would not fall. Officials described the operations as intended to avert yield contagion,7 which means preventing one bad auction from raising every other borrowing rate in the country. Mortgages, car loans, corporate debt and student loans are all priced from government bonds. When that yield moves, all of them move with it.

Separately, in January, the mortgage agencies had been directed to buy two hundred billion dollars of mortgage bonds, which performs the same function one market over.

Those three actions are better read as one behaviour than as three policies. An authority is buying its own long-term debt to hold the price up. It is funding that purchase by borrowing short. And it has instructed a second institution to support the adjacent market. The shape has a name: managing the yield curve. The novel part is not that it is happening, since central banks have done it before. It is that the central bank is not the one doing it. The borrower is.

This is the seigniorage. Not a printing press, not a crisis, nothing you could point to on a given morning and call an event. Just a sustained, technical, entirely well-mannered decision that the holders of long-dated money will absorb an arithmetic problem nobody wants to solve inside a budget.

This matters, rather than being technical housekeeping, because of what it does to the meaning of a government bond. Such a bond has two properties: it pays a known amount, and its price is set by people who are free to refuse it. The first is intact. The second is what is being renegotiated, and the second is the property that made it the reference point for the price of everything else.

There is also a blunt arithmetic reason the practice will not stop. At current rates, every additional trillion borrowed adds roughly fifty billion a year in interest, permanently. That arithmetic only resolves if the rate falls, and the rate will not fall on its own while the borrowing continues. What remains is managing it. An institution that begins managing a price because the alternative is unaffordable rarely finds a convenient moment to stop.

Section three

What the bond market is saying

A bond yield has two parts inside it, and they mean entirely different things. Separating them is the single most useful thing anyone can do with a chart of interest rates.

The nominal yield is the headline number, what the bond pays in dollars. Subtract the market's inflation forecast from it and what remains is the real yield, the true cost of borrowing. If the nominal yield rises because people expect more inflation, nothing fundamental has changed. If it rises because the real yield rose, lending to that borrower has genuinely become more expensive.

In 2026 almost all of the move was real. The ten-year rose about forty-eight basis points. The real yield inside it rose forty-one. The inflation expectation rose seven. This was not an inflation scare. It was a repricing of what it costs to lend to the United States government, which is a different and more serious event.

The shape of the yield curve says the same thing from another direction. Short-term rates are set by the central bank. Long-term rates are set by whoever is willing to hold thirty years of exposure to a borrower. When the gap between them widens while the central bank sits still, investors are charging more for time, and they are doing it without raising their inflation forecast. The thirty-year sits at 5.18 percent after touching 5.33, the highest since 2007.8

One detail does not fit any of this, and it is the one worth sitting with. The market's expectation of how much bond prices will move has fallen to a one-year low.9

Volatility, in this context, is not how much prices have moved. It is how much traders are paying to be protected against future movement, which makes it a measure of expected surprise. Normally it rises when the situation is unstable, since protection becomes more valuable when the ground is shifting.

Here the ground is plainly shifting. The thirty-year yield is fourteen basis points from a nineteen-year high, the budget deficit is the largest in peacetime memory, and the instrument that measures expected surprise in the bond market is asleep.

Volatility does not fall because a risk resolved. It falls because something is standing on it.

In this instance the something has a name and a published operating calendar: an authority buying long bonds specifically so the price does not move as it otherwise would. Suppressed volatility is not the absence of risk. It is risk with a lid on it, and the lid is a policy choice that can be revised.

The practical consequence is that insurance against a disorderly bond market is currently cheap, and it is cheap for a bad reason. Cheap insurance sold for a bad reason is generally the most attractive thing available.

Section four

Two economies

American output is strong. The Atlanta Federal Reserve's running estimate of quarterly growth has been near four and a half percent, business investment is booming, and the main manufacturing survey is at a multi-year high.

American employment is not. July payrolls came in at minus twenty three thousand against an expected plus eighty three, with a hundred and three thousand knocked off the two previous months in revision.10 Wage growth is the slowest since 2021. Consumer sentiment sits at 51, a level normally seen in the middle of a recession.

Payrolls count the jobs added or lost across the economy in a month, and the figure is revised twice as better data arrives. The unemployment rate counts only people who are actively looking for work. Someone who stops looking leaves the calculation entirely, which is why the rate can hold steady while the situation deteriorates. It held at 4.1 percent in July, and it held there because people left the labour force, not because they found jobs.

Both facts are true simultaneously because the growth is arriving from one place. Capital spending on artificial intelligence runs somewhere between one and one and a half percent of the entire economy, and on some estimates it produced the majority of the first half's growth. A data centre is an enormous economic event and a very small employment event: concrete, steel, transformers, chips, and a few hundred permanent staff. The marginal unit of American output now employs almost nobody. That is not a criticism of it. It is a description of what has been built.

The disagreement is visible in the forecasts. Recession has a technical definition anchored in output, and by that definition the optimists are correct. Households are experiencing something closer to what the pessimists are measuring. Both readings are accurate, because they are measuring different economies that happen to share a border.

Section five

A central bank nobody can read

The Federal Reserve has held its policy rate at 3.50 to 3.75 percent all year.12 That reads as stability. It is not: it is the residue of an unresolved argument.

The policy rate is the one interest rate the central bank sets directly. Every other rate in the economy, from mortgages to corporate loans, is priced in relation to it. Twelve officials vote on where it goes, and their disagreements are published, which makes the committee unusually legible for an institution of its importance.

At the July meeting the vote was nine to three, and all three dissents were in favour of raising rates. Three officials wanted to make borrowing more expensive in the same month that payrolls turned negative. In June the committee's own projections shifted to imply higher rates by year end, with half the participants forecasting an increase. The chair was appointed in May and nobody has yet observed how he behaves through a full cycle.

Markets currently price roughly two-in-five odds of an increase in September, better than even odds of one by December, and no cuts at all this year. Whether that pricing is correct is, I think, the wrong question to ask about it. The remarkable fact is not the level. It is that a serious market cannot agree on the direction of the next move by the world's most important central bank, ninety days ahead.

The uncertainty exists because the central bank's two objectives have come apart. It is required to keep prices stable and employment high, and those normally point the same way: an overheating economy produces both inflation and jobs, so cooling it addresses both. Not now. Inflation on the measure the Fed actually targets has been stuck at 3.3 percent for four consecutive months,13 held up by services and healthcare rather than anything an interest rate reaches. Employment is deteriorating. No single interest rate addresses both, so the committee is not really arguing about the economy. It is arguing about which half of its job to fail at.

Add the institutional question and it worsens. A central bank under public pressure from the executive, with a new chair and a divided board, is not simply a forecasting problem. It is a credibility problem, and credibility problems get priced in the same place the first section pointed at. That, I think, is what the gold market is looking at. Not the next twenty five basis points, but who will be setting them in three years, and under what pressure.

This is the connective tissue of the whole piece. Fiscal arithmetic that only works at low yields, plus a central bank whose independence is being tested, is the exact configuration under which the tax gets levied on money rather than on income.

Section six

The bubble is the financing

The question asked about artificial intelligence is usually whether the revenue is real. That is now the wrong question, and answering it correctly leads to the wrong conclusion.

The revenue arrived. Run rates at the major labs went from single-digit billions at the end of last year to roughly a hundred and ten billion in aggregate. The hardware is genuinely scarce rather than rhetorically scarce: memory contract prices nearly doubled in a single quarter, the advanced packaging capacity needed to assemble the chips is booked out at the leading manufacturer, and the largest chip company guided to over a hundred billion dollars in a single quarter while assuming no revenue from China at all.14 Whatever this is, it is not the year 2000, when the leading companies had no earnings whatsoever.

The fragile part is how it is being paid for, and that has changed faster than the revenue did. In 2024 the large cloud companies issued about twenty billion dollars of bonds. In 2025, a hundred and nine. Estimates for all-in artificial-intelligence debt in 2026 run to five hundred and seventy billion, which is close to one dollar in seven of every high-grade corporate bond sold in America this year.

Around that sits a second structure that does not appear on the same balance sheets.

Special purpose vehicles are separate companies created to hold an asset and its debt, so that the borrowing does not appear on the parent's accounts. Private credit is lending by investment funds rather than banks, which involves less disclosure and no public price. And vendor financing is an arrangement where the seller lends the buyer the money to make the purchase.

That last one deserves attention. When a chip supplier underwrites the obligations of the customer buying its chips, the two are no longer independent parties. The risk that appeared to be spread across a supply chain turns out to be one risk wearing several coats, and it becomes visible as one risk only when something goes wrong.

The comparison people reach for is the year 2000, and it is the wrong one, though the right one is standing next to it. Technology shares in 2000 traded at sixty times expected earnings with no profits underneath. Today's leaders trade near twenty three times with fifty percent margins. That comparison does not hold. What does hold is the fibre-optic cable laid alongside those shares: financed with debt, in quantities that proved correct about the internet and catastrophic for the companies that borrowed to build it. The cable stayed in the ground. It carries everything we do now. The financiers did not survive to see it.

So the position is two-tier. Own the physical bottlenecks, which are sold out and whose scarcity survives a financing accident. Stay away from anything whose equity depends on a levered intermediary successfully refinancing in 2027.

Section seven

What the money is buying

Follow the capital spending down the chain and it stops being a story about software almost immediately.

A model needs chips. Chips need advanced packaging, which is sold out, and high-bandwidth memory, where contract prices roughly doubled in a quarter and producers are earning margins normally seen only in a shortage. Those chips go into buildings, and buildings need electricity, which is where the story stops being about an industry and becomes about physical infrastructure.

The largest electricity capacity market in the United States has now cleared at its regulatory price ceiling three years running and still fell thousands of megawatts short of what it needs.15 One large cloud provider has disclosed roughly eighty billion dollars of orders it cannot fulfil for lack of power. A heavy-duty gas turbine ordered today arrives around 2031, not because of any shortage of money, but because three companies on earth make them and their books are full.

This is why I find the bottleneck tier more interesting than the model tier. If you believe the buildout continues, the constraint is where the pricing power sits. If you believe it stops, the constraint is where the residual value sits, because turbines and transmission and reactors do not become worthless when a software cycle turns. Both branches point at the same physical assets, and that is a rare thing to be able to say about anything.

There is also a knock-on effect that is only beginning to appear. Memory is a shared input. When demand from data centres takes the supply and doubles the price, every phone, laptop and car built in 2027 carries a higher cost of materials, and none of those manufacturers can charge more to cover it. That is a transfer of profit out of consumer electronics and into the three companies that make memory, and almost nobody is modelling it as an artificial-intelligence effect, because it does not look like one.

Section eight

Nobody has any buffers

This is the observation I did not go looking for, and the one I would keep if I had to throw the rest away. Five unrelated systems have run their shock absorbers down to nothing, at the same time, for entirely different reasons.

A buffer, in this context, is spare capacity that absorbs a surprise without anyone having to make a decision. A cash reserve, an emergency oil stockpile, unused production capacity. Buffers are invisible when they are full and are the only thing that matters when they are not.

The overnight facility that absorbed every money-market shock since 2021 holds four hundred and fifty six million dollars.16 Not billion. It held more than two trillion two years ago. At the same time the Treasury is scheduled to raise its own cash balance to a record just above one trillion by late October, drawn out of a banking system whose reserves are already falling.

Professional investors hold three and a half percent of their portfolios in cash, a level the survey that tracks it treats as a contrarian warning, because it means almost nobody has money set aside to buy a decline. Corporate borrowing costs sit within thirty basis points of the lowest premium over government debt ever recorded,17 meaning the compensation for taking corporate risk is as thin as it has been in the history of the measurement. The American emergency oil stockpile is at the bottom of its operating band.18 Spare oil production capacity across OPEC is around one million barrels a day, against eight million barrels a day of Gulf supply still offline from a war that has not ended.

Not one of those facts causes any of the others. A money-market facility and a petroleum reserve have nothing whatsoever to do with each other. That is precisely why it matters: there is no reason for a system to be thin in five unrelated places at once, and no mechanism that would coordinate a recovery across them either. Each has to be rebuilt separately, slowly, by a different institution with different incentives.

A shock does not have to be large. It has to arrive.

Section nine

The rest of the world

It is easy to write all of this as an American story. It is not, and two of the four largest pressures on US yields are being generated somewhere else entirely.

Japan is the one almost nobody watches and probably the most important. After thirty years of paying nothing, Japanese government bonds now yield close to three percent over ten years, the highest since 1996. Japan is the largest foreign holder of US government debt, and it holds it for a simple reason: for decades there was nothing worth buying at home. That reason is disappearing. Each step the Bank of Japan takes toward normal interest rates makes domestic bonds a real alternative for Japanese pension funds and insurers, and the largest foreign buyer of American debt takes its money home. Foreign holdings of Treasuries fell seventy two billion dollars in June, led by Japan.19

China is the opposite problem, and the reason global goods prices have stayed calm. It is in the thirty-seventh month of falling property prices, consumer inflation is a rounding error above zero, and house prices adjusted for inflation are below where they stood in 2006. An economy in that condition exports deflation: it makes goods more cheaply than anyone else and sells them into every other market, holding prices down abroad. That has been the anchor keeping American goods inflation near zero even through a tariff regime. If Beijing ever reflates at genuine scale, the anchor lifts, and every developed-market inflation forecast built on cheap Chinese goods needs rewriting.

Europe is quietly the good news, which almost nobody is positioned for. Its central bank has been raising rates because growth is holding rather than because inflation is threatening, German fiscal policy has turned expansionary, the equity indices are at records, and the banks are earning more than at any point in fifteen years. One detail is worth noting: Italian government bonds now yield the same as French ones. The country everyone worried about for a decade has converged with the country nobody worried about, which indicates that political risk in Europe has relocated rather than disappeared.

India is the structural one. It is the only major market down on the year, after foreign investors withdrew a record fifty two billion dollars in the first half. What makes it interesting is what happened next: domestic savers absorbed the entire outflow through monthly investment plans, and the market fell only single digits. That is a market becoming self-funding in real time, which changes what a foreign exit is capable of doing to it.

Section ten

What is actually cheap

If the argument above holds, the interesting positions are not bets on the direction of the economy. They are ways of owning the regime itself, which is a different shape of proposition and a more forgiving one, because it does not require the timing to be right.

Inflation-linked bonds. These are government bonds whose value rises with the official inflation index, so they pay a return specified in real purchasing power rather than in dollars. The ten-year version currently offers 2.34 percent above inflation, which is in the top tenth of everything on offer since 2008. The protection attached is close to free: the market is charging the equivalent of 2.32 percent expected inflation for it, during an energy war, with headline inflation at 3.7 percent and a central bank whose independence is being questioned.

A steeper yield curve. This position profits when long-term interest rates rise relative to short-term ones. Its appeal is that it wins for opposite reasons. If the economy weakens and the central bank cuts, short rates fall and the gap widens. If the fiscal problem deepens, long rates rise and the gap widens. It loses in only one world, the one where the government's arithmetic quietly resolves itself, and nothing in the data supports that branch.

The gold complex, bought on weakness rather than strength. Gold itself carries heavy speculative positioning after a thirty-three percent year, which argues for patience rather than avoidance. The more interesting expressions are silver, which is down on the year despite a sixth consecutive year of consuming more than is mined, and the mining companies, which earn a margin on every ounce and are currently earning record ones without having been repriced for it.

Volatility, while it is being suppressed. Equity volatility at 14.5, bond volatility at a one-year low, and oil options priced as though a contested shipping strait were a settled one. Insurance is cheapest exactly when the event it insures against feels least likely. That is not a paradox; it is the definition of how insurance is priced.

And the thing to avoid: ordinary corporate bonds. Thirty basis points of potential gain against a hundred and fifty of potential loss, with the weakest tier of borrowers already deteriorating underneath an index that still looks pristine. That is not a forecast. It is arithmetic about the shape of the payoff.

Section eleven

Six ways this goes

Two questions decide most of it. Does inflation cool enough to let the central bank ease, and does the credit financing the AI buildout hold together. Everything else is detail hanging off those two.

The most likely single path is the boring one, at roughly thirty percent: nominal growth grinds on, the Fed does at most one grudging increase, and asset prices rise in money that is quietly worth a little less each year.

What matters more than the ranking is which positions survive across the branches, and this is where the exercise earns its keep. A steepening curve wins in five of the six. Inflation-linked bonds win in the same five. Gold wins in four, and fails outright only in the eight percent branch where the war actually ends and the fiscal picture improves at the same time.

Equity direction wins in three and loses in three. That is not a call, and I think pretending otherwise is where most market writing goes wrong. It is an argument for staying invested and paying for insurance, which happens to be the cheapest thing available.

Section twelve

What would prove it wrong

I would rather be corrected early than defended late, so here is the list, with the thresholds attached.

• Gold below four thousand dollars for a month, with a credibly hawkish central bank. That is the regime claim failing directly. If the metal cannot hold its level while the institution reasserts itself, the level was about something smaller than I think.
• Payrolls back above a hundred and fifty thousand with upward revisions. The two-economies split closes, the household weakness was noise, and the case for owning the front end of the bond market goes with it.
• The November refunding raises coupon sizes and the long end absorbs it without complaint. Then demand for US duration is deeper than I judged, the buybacks were housekeeping rather than management, and the steepener is wrong.
• The AI credit rolls cleanly through 2027. Clean revenue recognition, contracts repricing upward, no distress in the levered middle. That retires the credit thesis entirely and the right response is to own the thing I was avoiding.
• Inflation expectations falling while gold falls with them. That would mean the market genuinely believes in the institution, and the premium I have spent this whole piece describing was never there.

There is also a boundary this page cannot cross, which I would rather state than have found. The data here closes on the morning of 28 August 2026, hours before the new central bank chair gave his first major address. Every rate, currency and gold view above was formed without knowing what he said.

In summary

What this adds up to

Six conclusions, ordered by how much confidence I have in them rather than by how interesting they are. The first two I would defend firmly. The last two are forecasts, and forecasts about credit are wrong far more often than the people making them admit.

1. Observable fact The United States has begun managing the price of its own debt.

   It froze the size of its long-term bond auctions, moved its borrowing to short-term IOUs, doubled its purchases of its own outstanding bonds, and directed the mortgage agencies to buy two hundred billion dollars of mortgage debt. Those operations are published and running on a calendar. Nothing here is inferred.

2. High confidence Output and employment have separated, and the aggregate figures no longer describe a household.

   Growth near four and a half percent alongside negative payrolls, the slowest wage growth since 2021 and consumer sentiment at 51. Both readings are correct because the growth is arriving through capital spending that employs almost nobody. A strong growth number has stopped being evidence that conditions are improving for most people.

3. High confidence The system is thin in five unrelated places at once.

   An overnight facility holding four hundred and fifty six million dollars rather than trillions, professional investors with three and a half percent cash, corporate borrowing costs near record-low premiums, an emergency oil reserve at the bottom of its band, and spare oil production of about one million barrels a day. None causes any other. Thin buffers do not predict a shock; they decide how far one travels.

4. Medium-high Gold's move is about the issuer, not about inflation.

   The bond market's own inflation forecast rose seven basis points on the year while gold rose thirty-three percent and central banks moved their reserves. If this were an inflation trade it would appear in inflation instruments. It does not. The qualification is real, though: within 2026 gold still moved with real interest rates, so the regime claim is about the level rather than the swings, and a genuine spike in real rates would still hurt.

5. Medium, and a forecast The fragile part of the AI buildout is the financing, not the demand.

   The revenue arrived and the hardware is genuinely scarce. What changed faster is how it is paid for: roughly five hundred and seventy billion dollars of AI-linked debt this year, plus separate structures that do not appear on the same balance sheets. The honest caveat is that the market currently disagrees with me. That paper keeps clearing and the levered middle rallied through August. This is a claim about 2027, and being early is indistinguishable from being wrong until it is not.

6. Medium, and a judgment Insurance is cheap in several markets at once, for a bad reason.

   Inflation-linked bonds paying 2.34 percent above inflation with the protection attached costing almost nothing. Bond volatility at a one-year low while the thirty-year yield sits near a nineteen-year high, because an interested party is holding it down. Equity volatility at 14.5. Each is cheap because the risk it covers is currently invisible, which is not the same as the risk being small.

If I had to compress all of it into one sentence: the safest assets in the system are the ones quietly carrying the bill, and almost nothing in their price says so.

That is not a prediction of a crash, and nothing above requires one. The mechanism described here works slowly and can continue working for years. It is a statement about which risks are being paid for and which are being absorbed silently, and about the fact that the compensation for holding the safe thing has quietly gone negative in real terms while its credit rating stayed exactly where it was.

In closing

Five open questions

The most useful output of all this was not the conclusions. It was the short list of things I could not settle from public data, each of which would reprice several markets at once.

1. Is the Fed's next move up or down? The market prices both within ninety days of each other. Nothing else moves every asset class simultaneously.
2. Does the AI buildout get refinanced, or repriced? Real revenue, largely unaudited, against half a trillion of debt that has to roll.
3. Does the war end, or escalate? Brent at eighty eight sits between the sixties and a hundred and thirty, and both tails are live in the same quarter.
4. Is the United States actually in fiscal dominance? The machinery is assembled and operating. Bond market pricing still assumes the old rules apply.
5. Is China's deflation terminal, or cyclical? The largest slow variable in the system, and the anchor under everyone else's goods prices.

Seigniorage was always the quietest tax a sovereign levied. The work is just to notice who is paying it now, and in what currency.