To Save the Economy, We May Have to Break It: Why the Fed's Hike Threatens History's Biggest Buildout

Last week the Fed raised rates for the first time in three years and the vote was a clean 12 to 0. Just seven weeks earlier the same committee had split 9 to 3 in July, with nine officials wanting to hold steady and three pushing to hike — meaning three-quarters opposed tightening at that moment. By September no one dissented. What changed was summed up by Fed Chair Kevin Warsh at the press conference: summer inflation prints showed price growth was not fading on its own and the current stance was not doing enough to curb excess demand. That single line gave the bond market the confirmation it was waiting for. The 10-year Treasury yield, the reference rate that prices every other rate on earth, closed above 5 percent, its highest finish since 2007.

From split to unanimous in seven weeks

History makes the unanimity harder to read as harmless. Since 1965 the Fed has steered 11 tightening cycles. The smallest lifted rates by 1.75 percentage points, the largest by 13. The typical cycle lasted about a year and a half and added roughly 4 points. Only one of those 11 stopped after a single hike: in March 1997 Alan Greenspan raised once, looked around, called it enough and cut 18 months later. The Fed's current projection rejects that one-and-done story; 16 of 18 officials expect at least one more increase before year end. A day before the meeting the Conference Board's chief U.S. economist went further, headlining a brief "More inflation is coming. The Fed rarely stops after one" and penciling three consecutive hikes in September, October and December.

Counting recessions sharpens the point. New York Fed researchers tallied 14 tightening cycles between 1955 and 2009. In 10 of them a recession followed within 18 months of the final hike. Of the four that did not, one produced a jump in joblessness and a credit crunch that many economists, including Nobel laureate Milton Friedman, treat as a recession, leaving just three clean soft landings. Three times in half a century the United States raised rates without breaking the economy. The question is therefore not how big the move is, but where the cycle settles once it gets going. Like opening a sluice, the water erodes wherever it flows.

Why soft landings are so rare

For the best part of 70 years the transmission ran through one visible door: housing. The chain works in three steps: 1) mortgage rates rise, 2) buying and refinancing turn expensive, 3) sales and construction cool, households spend less and the economy slows. Since the end of World War Two housing has been the Fed's most direct brake. That brake is now worn. Only about 40 percent of U.S. households carry a mortgage and of those 78 percent are locked below 6 percent, two-thirds below 5 percent and half below 4 percent. If you sit on a 4 percent loan while new loans price near 7 percent, you simply do not sell. The Fed's July report to Congress labels this rate lock. The sales data shows it: existing homes sell at roughly 4 million a year against a 15-year average near 5 million, and last year was the slowest since 1995. Housing is already frozen; pushing rates higher now presses a pedal with little pad left.

If the favorite brake no longer grips, where do the hikes settle? Warsh answered himself 19 days before the hike: enormous and growing sums of capital are pouring into AI-related infrastructure of every kind. That line pointed straight at data centers. JPMorgan estimates the global build at about 5 trillion dollars through 2030. A Columbia finance professor did the math just for the U.S. leg: if built as planned it would consume roughly 2.8 percent of annual U.S. output every year through 2032. For comparison, building every railroad in the 19th century took about 2.5 percent of the economy, electrifying the whole country about 1 percent and building the interstate highway system about 1.5 percent. This is the largest project the United States has ever built, larger than the railroads.

History's biggest construction site

The builders are the most profitable companies in history, yet the funding does not come from the till. Operating cash flow, equity sales and all internal sources cover only about a quarter of the 5.5 trillion dollar need. JPMorgan expects the other roughly 4 trillion to be borrowed. That makes the price of money not a footnote but the whole plot. Issuance betrays it: from 2020 through 2024 the five core tech firms issued about 35 billion dollars a year in bonds; in 2025 the figure reached 93 billion and in the first seven months of this year alone it hit 132 billion. Consultant Chatham Financial puts AI infrastructure spending this year above 830 billion dollars — about half the entire U.S. investment-grade bond market, two-thirds of leveraged loans and more than the whole high-yield market combined. Vanguard expects these firms to spend more than 1 trillion dollars a year in capital outlays for the next three years. This is not a one-off check but a standing appointment with the bond market for the rest of the decade, and each visit will cost more than the last.

The market is already writing that price. A spread is the extra yield a company pays over the government, a simple fear gauge. This year the spread for those five names widened by about 30 basis points while the broad investment-grade market moved just 2 — fifty times the move, from some of the safest credits on earth. Move down the stack and the tone sours. Debt that funds the actual data center projects prices about 100 basis points wider than the parents' own bonds and by another 200 as you reach high-risk territory. At the bottom repricing is happening in real time while the build continues. CoreWeave, which rents graphics processors to others, had a 2.6 billion dollar loan pulled mid-syndication; the coupon rose by a full point and terms tightened to full amortization over the life, a coverage test that incoming cash must beat outgoing debt by a set margin plus a minimum cash balance. It ultimately funded at 10.4 percent. Within a year CoreWeave's quarterly interest bill rose from 267 million dollars to 640 million; in the second quarter interest alone consumed 42 cents of every dollar of adjusted earnings.

The fear gauge: how spreads tell the story

At what price does the build stop paying for itself? Compute-pricing firm Mercatus modeled a single graphics processor fleet at different costs of capital. At 6 percent the return is comfortable, at 10 percent barely breakeven and at 12 percent the operation sinks outright, with the breakeven around 11 percent. Mercatus also sorts buyers by their cost of capital: entrenched firms with strong balance sheets pay 6 to 8 percent on new deals, late-stage AI firms with solid revenue pay 10 to 14 percent and early-stage venture-backed AI firms pay 15 to 20 percent. That puts the top of the market above breakeven, the middle exactly on the line and the bottom already underwater. Picture a three-story building: the top floor stays dry, the middle sits at the waterline and the basement is flooded.

Borrowing cost also behaves like rent: it resets, and it rises two ways. The quick way is floating-rate debt at the bottom of the stack. That CoreWeave loan prices off the overnight rate that moves with the Fed plus five and a half points, roughly 10.5 percent all-in after last week's hike, so the loan grew dearer before the next payment arrived. The slow, structural way is through refinancing at the top. Bonds at the top are fixed, but every loan matures and every bond comes due and almost no one repays outright; they roll the debt at that morning's market price. A firm that borrowed cheaply years ago will not get the same price when it returns for new debt. The median tightening cycle since 1983 added more than three percentage points, 16 of 18 officials expect at least one more rise this year and the Conference Board expects three. The industry does not borrow once; each return will be pricier.

Why borrowing works like rent

That is why this cycle differs from every one before it. In the past a hike mainly hit demand: mortgages, card balances, auto loans — the household shopping list. This time the biggest hit lands on supply, on the companies carrying towering debt while trying to build the future. The common mistake about rate hikes is to look for damage on announcement day. For most borrowers nothing breaks that day; it breaks later, when someone must come back for more money and that money costs more than last time. We are watching enormous debt demand and a hiking cycle that has only just begun run at each other at full speed, like an Oklahoma drill.

Imagine letting AI firms burn and the pain stays contained. Apollo's chief economist shows it does not. He tracked data center spend as a share of the economy: it is currently growing by about 0.85 percentage points of GDP a year. Housing's fastest boom phase from 2002 to 2005 grew by half a point a year; telecom expansions in the 1990s peaked at 0.15. The AI build is rising almost twice as fast as the housing bubble, and we remember how that ended. AI is not just fast; it is the engine of growth. This year AI capital spending will contribute roughly 1.5 percent to U.S. economic growth. Private data center construction now runs at 75 billion dollars a year, up 57 percent in one year and more than double in two, and Census data show the country now spends more on data centers than on traditional office buildings. The good news is nothing has collapsed yet. No one is pulling back, nothing has stopped moving. The bad news is absence of damage is not proof of safety. JPMorgan went back 20 years to see what the five biggest tech firms do with each dollar of revenue: for most of that time capital spending never exceeded 13 cents on the dollar; this year it reaches 41 cents and for the first time the chart shows free cash flow turning negative. The hole to be funded by debt will grow every year, and with it the interest bill to fund it. That is the real picture behind the latest decision.