Japan's Treasury dump and AI's power problem: two clocks on the US economy

The story that made cable news was friendship. The story that moved bond markets was a handwritten to-do list, a pot of Treasury money that does not need Congress, and the largest foreign holder of US government debt selling paper into a market already at nineteen-year highs. Meanwhile, the sector carrying more than half of recent GDP growth is betting on power infrastructure with an eight-year wait list.

The receipt was a doodle

At 11:33 a.m. on Friday, July 31, a photographer with a long lens captured something sitting on the desk of the US Secretary of the Treasury: a handwritten note with one line legible from across the room — “Buy Japanese yen, 5 to 10 billion dollars.” Hours later, for the first time in twenty-eight years, the United States walked into the currency market and bought yen. The official receipt was a doodle on a notepad.

President Trump, asked aboard Air Force One why America was supporting the yen, offered friendship, financial strength, and an aside about Pearl Harbor. Tokyo's finance ministry answered in stiffer language: the operation countered excessive volatility and disorderly movements in the yen, taken in coordination with the US Treasury. Translated out of Treasury speak, the yen was falling, Japan was hurting, and Washington stepped in to make the pain stop.

The timing mattered. US thirty-year borrowing costs had just hit their highest level since July 2007. And the country America was helping holds more US government debt than any other foreign nation on Earth. So the polite question behind the polite answers is sharper: what was America actually buying?

The Secretary's piggy bank

Governments do not wire-transfer yen from a checking account on a Friday afternoon. The money came from the Exchange Stabilization Fund (ESF) — a pool at the Treasury the Secretary can spend without asking Congress. He needs the president's signature. That is it. The statute says his decisions are final and, quoting the law itself, may not be reviewed by another officer or employee of the government.

The ESF was created to stabilize the exchange value of the dollar. History suggests “stabilize” was never as clearly defined as it should have been. Past Secretaries reportedly used it for an Oriental carpet, a forty-two-person trip to London, Moscow, and Mexico City with no identifiable official purpose, and whatever else the Secretary and president agreed counted as stabilization. Think of it as a discretionary FX credit card with no itemized receipt requirement.

The detail that changes the story is what was inside the fund. America did not buy those yen with dollars. It sold euros — roughly $13 billion of euro assets held in the ESF went out the door. The European Central Bank learned after the trades had cleared. Senior European officials told the Financial Times the move broke decades of convention and had never happened before. When asked, Treasury said it does not coordinate with anyone on how it allocates the fund — true, and also the fiscal equivalent of pleading the fifth.

One more date on the timeline: on July 23, Treasury published its foreign exchange monitoring list — the official watch list for trading partners suspected of currency manipulation. Japan was on it. Eight days later, the same building that released the report bought Japan's currency. America broke a twenty-eight-year rule, paid by dumping Europe's currency, told Europe afterward, and did it for a country it had just flagged as a currency risk.

Japan's doom loop

Japan owes more relative to GDP than any major country on Earth. Servicing that debt jumped nearly 11% this year to roughly 13 trillion yen against a total budget of 122.3 trillion. Right as the debt bill spirals, the yen has collapsed toward forty-year lows. The mechanism is basic math, not sentiment.

For decades Japan pinned borrowing costs near zero. It recently started rolling that back. Japan's benchmark rate now sits around 1% while America's sits between 3.5% and 3.75%. Capital does not catch feelings. It goes where it gets paid the most. Investors dump yen, chase dollar yield, and the currency falls.

Armchair economists ask why Japan does not just raise rates and close the gap. Because it cannot — not without detonating the balance sheet it built to survive zero rates. To hold borrowing costs at zero for that long, the Bank of Japan bought its own government's bonds until it owned roughly half the entire JGB market. Those bonds pay almost nothing because paying nothing was the point. But commercial bank reserves at the BOJ earn whatever the policy rate is. Raise rates to save the yen and the central bank starts paying out trillions in interest while collecting almost nothing on the pile it is stuck holding. Japan can defend the yen or it can afford the debt. Not both.

The FX intervention feedback loop

Currencies trade in pairs. To push the yen up, you buy yen and sell the other leg — almost always dollars, since the dollar sits on one side of nearly 90% of global FX trades. Japan has the dollars. Roughly $1.3 trillion in reserves, with about $930 billion in securities — mostly US Treasuries. A bond is not money. You cannot hand a Treasury note to a currency dealer and ask for change.

So Japan has two options: sell bonds or pawn them. It chose to sell. In one month Tokyo spent roughly 11.7 trillion yen — nearly $74 billion — on the largest currency defense in its history. In May, Japan's Treasury holdings fell by $66.7 billion, the single largest one-month drop by any country in recorded history.

It worked for about three weeks. Then the feedback loop kicked in. Bond prices and yields move in opposite directions. Flood the market with Treasuries nobody asked for and yields rise. Japan sells US debt, American borrowing costs go up, the rate gap widens, and capital still flows to the higher yield. Every dollar Tokyo raised to push the yen up made the reason the yen was falling slightly worse. That is the FX intervention doom loop — a machine that eats its own tail.

The pawn shop nobody was using

When your largest foreign lender demonstrates in public that it will sell your paper to defend its own currency, you build a second way to raise dollars. The Federal Reserve operates one: the Foreign and International Monetary Authorities (FIMA) repo facility.

Here is how it works. A foreign central bank — say, the Bank of Japan — hands $100 in US bonds to the Fed. The Fed gives $100 in cash. The next day the central bank buys the bonds back for $101. The $1 spread is the Fed's interest. The whole point is to give foreign countries quick dollar access so they are not forced to dump Treasuries on the open market. It is what you build for a creditor you cannot afford to lose.

Forty-eight hours after the July 31 joint intervention, Treasury Secretary Scott Bessent posted on social media that the FIMA repo facility is an important backstop and that he would encourage the Fed to expand it in the coming months. The strange part: nobody was using it. The facility has a $60 billion per-country ceiling. The balance was zero. Bloomberg reported it had gone unused for an eighth straight week.

On CNBC, the host laid it out directly: one thing we definitely do not want is Japan selling Treasuries to defend the yen. Bessent responded that the goal was to protect the US economy and keep volatility offshore, and that the bond market was much smaller when FIMA was created in 2020 — so upsizing would be reasonable. He did not mention the facility was empty, or that its $60 billion ceiling is less than the $73 billion Japan spent in one month alone. You do not publicly ask the central bank to enlarge an unused facility unless you expect it to get used.

Mark Sobel — who spent roughly four decades inside Treasury on international monetary policy — wrote that aside from supporting an ally, Treasury may have been concerned that a weak yen could push American long-term rates higher. The president says friendship. The finance minister says stability. The career Treasury hand says there is more to the story. The public record agrees with the third version.

The other bet: AI on a grid that moves at Blockbuster speed

Creditor stress on the liability side of America's balance sheet is one structural risk. Growth concentration on the asset side is another. If Japan's Treasury sales threaten the cost of money, the sector currently carrying the economy threatens the return on money.

In 2024 the largest tech companies spent a little over $200 billion on AI infrastructure. In 2025 that figure approached $400 billion. Their own estimates for 2026 exceed $700 billion — nearly doubling every year. Deutsche Bank's George Saravelos wrote to clients that AI-related spending appears to be saving the US economy right now; without it, the US would be close to or in recession. Last quarter AI accounted for more than half of real GDP growth. The concentration is not subtle.

Every bubble take you have heard argues the wrong question — whether chatbots work, whether AI revolutionizes every industry, whether a model that miscounts letters in “strawberry” can take your job. The real bet is narrower and scarier: can the physical infrastructure arrive fast enough to justify what investors are paying today? Money is disappearing into AI faster than returns are showing up. Bubbles do not pop when technology fails. They pop when capital gets tired of waiting.

Clock one: the power clock

Wall Street loves the Mark Twain advice: do not bet on miners, buy picks and shovels. For AI that means data centers, power, transmission. Nobody mentions America is running out of picks and shovels.

A data center does not plug into the grid and ask nicely. It joins an interconnection queue — the wait list for grid access. In America's largest power market, that wait stretched from under two years in 2008 to over eight years today. Data centers already consume more than 4% of US electricity and are projected to reach roughly a fifth of consumption by 2035. They run flat out around the clock, not like office towers that sleep on weekends.

The workhorse behind much of that buildout is the gas turbine — roughly 40% of US electricity, among the most complex machines on Earth, built by only a handful of companies. Three manufacturers control over three-quarters of the global market. GE Vernova's order backlog ballooned from roughly 80 GW to 116 GW; its CEO says the company will be largely sold out through 2028. In 2024 alone developers ordered nearly three times more turbine capacity than every factory on Earth could produce in a year. Transformers now take over two and a half years with an estimated 30% supply shortage. New transmission lines average six and a half years to complete.

Tech moves at Netflix speed. The grid moves at Blockbuster speed. Developers have paid $25 million not for a turbine but to reserve a place in line. Others have bolted retired jet engines to the ground for makeshift power. Money already ran the experiment of throwing cash at the problem. Money lost.

PJM: when the auction hits the ceiling

PJM Interconnection serves roughly one in five Americans across thirteen states from Chicago to Washington, DC. Its biggest annual event is the capacity auction — a billion-dollar pinky promise that power plants will be ready on the worst days of the year. For most of this century that insurance was cheap because demand barely grew. In the auction covering 2024, a promise of power cleared at about $29 per megawatt-day.

Then data centers arrived by the dozen — each swallowing as much electricity as a small city, demanding power now, guaranteed, forever. The next auction cleared near $270 — an more than 830% spike. The regional bill went from $2.2 billion to $14.7 billion. Governors leaned on PJM. Federal regulators signed off on a legal ceiling. The next three auctions hit that ceiling anyway. What looked like a spike became the floor.

In December 2025 something happened that had never happened in PJM's history: the auction came up short. PJM went shopping with more money than ever, paid the maximum price allowed under federal law, and still could not buy enough supply to meet its reserve margin target. All that money cannot summon a gas turbine, transformer, or transmission line out of thin air when manufacturers are booked years out. The bill does not stay with Big Tech alone — it flows into household electric rates. Monitoring Analytics found that of $16.4 billion in recent auction costs, roughly $6.5 billion — forty cents of every capacity dollar — traces directly to data centers.

Clock two: the chip clock

Data centers are warehouses for the most expensive product in the buildout: AI chips. They have two problems. First, they do not age well. A cutting-edge AI chip stays cutting-edge for somewhere between two and six years depending on who you ask. The infrastructure serving it lasts decades. Chips age like smartphones; power equipment ages like bridges. When chips go obsolete, the grid assets built to feed them stay on utility books for decades and a chunk of the cost shifts to ratepayers.

Second, reported earnings may depend on a boring accounting guess: how long a server stays useful. Stretch the depreciation schedule and annual expenses fall; profits look bigger today. Michael Burry argues Big Tech is understating depreciation by roughly $176 billion between 2026 and 2028 by assuming six-year useful lives when the real number is lower. Nvidia pushes back. The tech companies cannot agree with each other: Amazon extended server life to six years in January 2024 — cutting depreciation by $3.2 billion and adding $2.5 billion to profit — then walked part of it back twelve months later citing the pace of AI development. Meta went the opposite direction, extending life to five and a half years and adding roughly a dollar to EPS in 2025.

Same silicon. Same period. Opposite conclusions. That is the race:

• Clock one — new power infrastructure can take five to ten years to build.
• Clock two — today's AI chips may only stay economically relevant for three to six years.

Goldman Sachs found that historically about 72% of data-center projects due within a year opened on schedule. Adjusting for today's supply chain, they expect about 60% next year and 50% two years out — flip a coin. Bloomberg's energy research ran two forecasts side by side: chip sales imply one buildout path; grid capacity implies another. By 2033 the gap is 63 GW — roughly half as much power as every operational data center in America today, on a grid with an eight-year wait list.

Success broke the grid. Success kills the chips. The better AI does, the more demand floods the grid and the longer clock one takes. The faster AI improves, the quicker hardware goes stale and the shorter clock two gets. Success winds both clocks against each other.

Two stories, one surface

None of this is hidden. It lives in Treasury data releases, FIMA balance sheets, PJM auction results, SEC depreciation footnotes, and occasionally a photograph of a notepad nobody was supposed to see. The headline story and the story underneath are rarely the same one.

On the rates side, America intervened to help an ally whose currency defense was becoming a Treasury supply problem — while thirty-year yields sat at their highest level since 2007 and the FIMA pawn shop sat empty with a capacity smaller than one month of Japan's FX spending. On the growth side, the economy's current engine is capital-intensive infrastructure with a delivery schedule the physical world cannot match and an accounting clock that may be shorter than the power clock.

For long-term equity investors, the implication is not panic. It is framing. Macro context changes which businesses can survive higher rates, which can pass through energy costs, and which are priced as if both clocks have already synchronized. Outpick underwrites individual companies against that backdrop; we do not trade FX interventions or capacity auctions. But ignoring the wiring underneath the headline is how portfolios get surprised by things that were public all along.

Frequently asked questions