Jeff Gundlach laid out the Fed's trap this week: if the Fed hikes, it worsens the interest expense problem, since so much borrowing sits at the short end; if it cuts, it worsens the inflation problem.
He's right about the trap. I think he's asking the wrong question. The real question is which way out Washington has already picked, and what breaks if that exit gets blocked.
The debt math leaves one exit
The federal deficit hit $2 trillion in the first 11 months of fiscal 2026, and the US now spends more on yearly interest than on national defense. When the deficit is bigger than the interest bill, every interest dollar is a borrowed dollar. The principal grows, the refinancing wall grows, and the liquidity needed to roll it grows with it.
Nobody pays that down. The only way out is to grow nominal GDP faster than the debt. Debt-to-GDP stops rising without a single dollar of principal repaid.
Step one: hand the balance sheet to the banks
The Fed restarted Treasury bill buying in December. Keep the size in perspective. From early January to early July, it bought nearly $250 billion in bills, about $90 billion of which was reinvested mortgage-bond runoff, and Fed assets rose about $150 billion. That's a bridge, not a flood.
The real handoff runs through the banks. The revised leverage rule for the largest banks took effect April 1, 2026. Its stated purpose is reducing disincentives for big banks to engage in low-risk activities such as Treasury market intermediation. My read: it also frees their balance sheets to lend, and they are. Bank loans were growing 6.8% year over year in March, putting roughly a trillion dollars of new credit within reach. Bank lending creates money that reaches the real economy faster than QE ever did.
The problem: the curve is bending the wrong way
Banks borrow short and lend long. They need a bull steepener, where short-term rates fall faster than long-term rates, widening the margin on every new loan. We have the opposite.
The FOMC voted 12-0 on September 16 to raise rates a quarter point to 3.75%–4%, its first hike since July 2023, and 16 of 18 participants expect another increase. The 10-year sits near 5.17%, close to its highest since June 2007, and the 30-year recently touched 5.501%, its highest since June 2004. The front end moved hardest. In late July the 2-year yielded 4.244%; today it's 4.899%. The curve flattened.
Strip out oil and inflation looks tame. Core CPI eased to 2.4% in August, the lowest since March 2021. This hike bought credibility with the bond market. It wasn't a response to broad inflation.
The pressure valve is oil
Iran has proposed a ceasefire of up to 60 days, a phased reopening of the Strait of Hormuz, and an end to the US blockade. A deal before the midterms would help the White House at the pump.
Don't take it on faith. Trump has ruled out lifting the blockade before Iran shows sufficient goodwill. Former negotiator Dennis Ross puts the odds of a pre-midterm deal at only 30%, though both sides have stronger reasons to deal before the vote than after. We've already seen one failure: a mid-June memorandum produced a fragile ceasefire that collapsed weeks later.
If a deal lands and crude falls, Warsh gets cover to pause, then reverse. The front end rallies, the curve bull steepens, the dollar weakens and gold runs. Falling rates, a falling dollar and falling oil all reduce the capital the world locks up in hedges. Released capital gets levered and lent. That's act one.
Act two: the productivity boom
Act two is a replay of Greenspan's mid-90s. Real GDP ran 4–5% while core CPI held near 2–2.5%, and the Nasdaq 100 rose more than 500% from the end of 1995 to the end of 1999. Greenspan bet that technology was lifting productivity and refused to fight an inflation wave that never came.
Warsh has made the same bet out loud. At Jackson Hole he argued that AI could increase the economy's efficiency, enabling it to expand without creating inflationary pressures.
The buildout backs him up. Hyperscaler capex is expected at around $646 billion in 2026, about 2% of US GDP. The Apollo program ran around 0.6% of GDP, and the Manhattan Project about 0.4%. Corporations are paying for this one, increasingly with borrowed money, and banks are the natural lenders for the next leg: power, grids, factories and robotics.
The 10th Man: what if act one never arrives?
Here's where I part from the consensus bulls. The playbook depends on rates falling. If the Strait deal fails and WTI clears $110, Warsh keeps hiking and the long end stays pinned above 5%. Markets already price nearly a 64% chance of another quarter-point hike in October. In that world, the AI buildout doesn't stop, but its funding costs go up and the weakest balance sheets break first.
Hyperscalers bend. They've shifted from funding capex out of cash flow to funding it in the bond market. The five hyperscalers averaged about $35 billion of debt a year from 2020 to 2024, issued $93 billion in 2025, and roughly $132 billion so far this year. The market already charges a premium: AI-related issuers trade around 115 basis points over Treasuries, versus 78 for the broader investment-grade market. And Goldman expects hyperscaler gross issuance to hit a record $420 billion next year. Higher-for-longer compounds that cost on every new tranche. Apollo's Torsten Slok warned that rising all-in yields could force the AI capex cycle to "self-throttle."
Not all hyperscalers carry the same risk. Microsoft, Alphabet, Amazon and Meta generate enormous cash flow and can absorb higher coupons. Oracle is the pressure point: it's the largest non-financial borrower in the Bloomberg high-grade index, and its 2054 notes yielded 7.8% in July.
Neoclouds break first. These are the companies that rent AI computing power, and they carry the most leverage in the stack. CoreWeave's Q2 interest expense reached $640 million, with debt-to-equity at 8.94 and net debt to EBITDA at 10.75. Higher yields raise both the interest cost on that debt and the discount rate applied to future cash flows. The credit market has noticed: CoreWeave's credit default swaps topped roughly 855 basis points in July, implying a 50% five-year default probability on a widely used pricing model. Strong operations don't cure a financing squeeze.
Data center developers inherit tenant risk. Many developers fund construction with bonds backed by long leases to a handful of tenants. Since June, high-yield data center spreads have widened, and questions about lease coverage, construction timelines, chip refresh cycles and tenant concentration have come back into focus. Concentration cuts both ways. Applied Digital's 400 MW North Ellen facility is fully contracted to CoreWeave, leaving it directly exposed to CoreWeave's creditworthiness. If the tenant wobbles, the landlord's bonds reprice.
Builders and equipment suppliers feel it last. Power gear, cooling, electrical contractors and construction firms sit on backlogs today. Backlogs lag. Rising funding costs first show up as deferred project starts, then as slower new orders a few quarters out. The transmission runs down the chain: hyperscaler capex guidance, then neocloud contracts, then developer leases, then equipment orders.
The counterpoint deserves airtime too. Capital Group's John Queen put it bluntly: if the hyperscalers see AI as a 30–50% return compounding investment, do they really care about 50 or 100 basis points on their spreads? For the cash-rich giants, probably not. For everyone financing the buildout with borrowed money, absolutely.
A failed act one delays the productivity story. It doesn't cancel it. But it changes who survives to collect.
What I'm watching
- WTI: below the recent rejection near $105–107 keeps act one alive. A clean break above $110 means no deal, sticky inflation and no cover for Warsh.
- The 2s10s curve: we need a bull steepener, not more flattening.
- The dollar and gold: confirmation that act one has started.
- AI credit spreads and neocloud CDS: the early warning system if act one fails.
- Q3 earnings capex guidance from the hyperscalers: any throttling shows up here first.
The bull case needs everything to go right in sequence. The bear case needs just one thing, oil, to go wrong. Position for the productivity wave, but size the leveraged links in the AI chain like the rate path isn't settled, because it isn't.
The numbers, not the headlines.
Live Loud!
The core macro framework comes from the piece you shared. If you know its author, consider a line of credit. I can also put this into a doc or file for publishing.