Every account of Kevin Warsh’s Jackson Hole speech on Friday told the same story: the Fed chair sounded hawkish, September hike odds jumped, semiconductors sold off. All true, all incomplete. What none of that coverage did was put the speech next to the specific balance sheets that spent this year borrowing on the assumption the next move was down. Hyperscalers and their affiliates have issued roughly $225 billion of bonds in 2026, close to a tenfold increase on last year’s pace. That paper was sold into an easing cycle. On Friday the man who runs the easing cycle said he is not sure there is one.
That is the story. Not the odds, the borrowers.
What Warsh Actually Said
The remarks themselves were less dramatic than the market reaction implies, which is what makes them interesting. Warsh recommitted to the 2 percent PCE target and set a bar: “We must be confident that underlying inflation is moving to our objective, clearly and at sufficient speed.” Then he closed the door on the summer’s good news. This year’s better than expected PCE and CPI prints, he said, “do not tell me that underlying trends have meaningfully improved.”
Markets did the arithmetic quickly. Odds of a quarter point hike at the September 16 meeting went from roughly 35 percent on Thursday to the high fifties by Friday’s close, and CNBC’s Monday analyst roundup put them at 60.4 percent. The two year yield jumped and the ten year added four basis points to 4.72 percent. The S&P 500 closed at 7,711.76, down 0.25 percent, and the Nasdaq gave up 0.52 percent to 26,402.42.
Worth remembering what the current setting is. The target range has sat at 3.50 to 3.75 percent since December 2025. A September move would not be a tightening cycle. It would be one hike into a market that has spent eighteen months pricing the opposite direction.
The Number Nobody Put Next to It
Here is the figure that should have run alongside every hike-odds chart on Friday.
J.P. Morgan estimates hyperscaler capital spending reaches $697 billion in 2026, and roughly a third of it is debt financed. Amazon, Alphabet, Meta and Oracle alone sold about $194 billion of bonds through early July, against roughly $108 billion across all of 2025. Fortune’s tally of hyperscalers and related issuers including Nvidia puts the 2026 figure at about $225 billion by midyear, a jump of more than 970 percent. Beyond that sits a financing gap of roughly $1.5 trillion between projected data center spending through 2028 and what these companies can fund from operating cash.
None of that is a problem at 3.50 percent and falling. All of it is a different conversation at 4 percent and rising.
The credit market started having that conversation before Warsh did. Median spreads on Amazon, Alphabet, Meta and Oracle paper maturing in two to four years widened to 40 basis points from 30 last year. On maturities past twenty years, spreads reached 118 basis points against 108.5. Those are small numbers on enormous principal, and they moved while the market still expected cuts.
Oracle Is the Stress Test
If you want to know what this looks like when it goes wrong, watch Oracle. Its total debt now exceeds $130 billion, sitting under about $248 billion in new lease commitments, and its five year credit default swap spread has widened roughly 310 percent to a sixteen year high in perceived risk. CNBC has been tracking the leverage question for weeks.
The industry-level ratio is the uncomfortable part. AI infrastructure generated something like $60 billion of revenue in 2025 against roughly $400 billion of capital expenditure. Borrowing against a gap that wide is a bet that revenue arrives before the debt service does. Higher rates do not break that bet. They shorten the amount of time available to win it.
A hike does not pop an AI bubble. It ends the era when the buildout could be financed as though capital were free, and it does so while revenue is still four hundred billion dollars behind the spending.
Friday gave a preview of the mechanism. Nvidia rallied nearly 9 percent on Thursday, then reversed more than 3 percent on Friday on no company news at all. The only input that changed was the Fed. When a stock that carries the index moves that hard on a rates signal rather than an earnings signal, the AI trade has quietly become a duration trade. We flagged the same transmission when the thirty year cleared 5.1 percent in May, and it is running faster now.
Where We Come Down
Warsh is right on the policy and wrong about the communication, and the two are connected.
He is right that a central bank should not declare victory on a couple of soft prints, and right to resist the pressure from an administration that wants cheaper money for its own reasons. Treasury Secretary Scott Bessent has been intervening directly to hold yields down, doubling long dated buybacks and describing it as signaling. A Fed chair who folds to that is not running an independent central bank.
But Warsh has also spent his tenure refusing to offer forward guidance, and Friday is the invoice for that choice. Because the market had no steer, it defaulted to the old assumption that rates fall, and roughly $225 billion of AI infrastructure debt got priced against that assumption. A single speech then repriced the whole complex in an afternoon. That volatility is not the market working. It is the market guessing, and guessing wrong for eight months.
The right move in September is the hike, and the thing Warsh owes alongside it is a plain statement of where the terminal rate sits. Not a dot plot he declines to participate in. A sentence. The companies building the physical layer of AI are making twenty year capital commitments, and they are currently doing it by reading his adjectives.
Watch the September 16 decision, but watch the bond calendar harder. The tell will be whether any hyperscaler tries to price large long dated paper in the two weeks before the meeting. If they rush to issue ahead of it, they have already concluded the cheap money is gone.