AI Debt Is Crowding Out Treasuries, and the US Economy Is Actually in Deflation
AI capex has grown big enough to push into the Treasury market and distort global trade and inflation data; strip out the AI-related parts and the US economy is mildly deflationary, while the Fed is hiking for the wrong reasons.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI capex can already move Treasury yields
Over the past 12 to 18 months, AI capex has consistently accounted for 30% to 70% of US GDP growth. More anomalous is that it has started to seep into the Treasury market: AI-related debt issuance is approaching $1 trillion, and debt's share of AI financing has risen from 15% to 20% a year ago to over 60%, making it the single largest block in both the investment-grade and high-yield markets. The result is that yield-sensitive investors, at equal duration, would rather hold hyperscaler debt than Treasuries, pushing the 10-year yield up. Kedrosky stresses this is not the market losing faith in America's ability to pay — the US can print money and will not default — it is competition at the margin.
— Paul KedroskyMore than half of global trade growth is AI
The WTO had forecast that global goods trade would fall noticeably because of US tariff policy; instead it grew about 4.5%, close to double the forecast. Digging into the data itself to find out why, the WTO found that over the past three quarters, 19% of global goods trade was AI-related, and close to 55% of global trade growth was AI-related — mostly NVIDIA GPUs. Kedrosky's conclusion: AI capex is masking the real impact of Trump's tariffs, because it is so large it drowns out every other signal.
— Paul KedroskyBoth sources of US inflation point back to AI
US inflation is still above 3%, yet consumers are more stretched and more indebted than they have been in a long time. Break inflation apart and the two biggest pieces are: first, AI capex pushing up energy prices, which in turn pushes up construction costs, materials and all kinds of services; second, the energy shock from the Iran war. Kedrosky counts the latter as AI-related too — he argues the Trump administration was emboldened by the success of the operation it ran in Venezuela using Anthropic, turning war into an app, and Iran is next. So both preconditions for inflation sit on top of AI.
— Paul KedroskyStrip out AI and the US is in deflation
If you remove from the US economy the inflation tied to AI capex and the energy inflation from the AI-emboldened Iran war, Kedrosky calculates the US is actually running about 0.25% deflation. That means the Fed is hiking for the wrong reasons — it thinks it is suppressing demand-side inflation, when the actual inflation is exogenous. This resembles the path into the 1929 crash: high rates stacked on a weak consumer are the classic precursor to a long balance sheet recession. He cites Richard Koo's work on Japan's lost decade and argues all the conditions are now in place.
— Paul KedroskyBalance sheet recession: ten years of paying down debt, no spending, no hiring
The hyperscalers will not go bankrupt, but they will spend a decade cleaning up their balance sheets — going from the least indebted companies in the market to the most indebted. A balance sheet recession is when companies, regardless of the level of rates, focus on deleveraging and getting debt off the books. The consequence: no investment in growth, no hiring, while at the same time looking for opportunities to cut headcount with AI. Companies like CoreWeave face a different order of problem — rolling over debt over the next five years, they may be forced into some form of insolvency. Kedrosky says this will be the most underrated macro event of the next five years.
— Paul KedroskyThe system is screaming: data center financing costs are rising on their own
The 10-year yield has risen about 100 basis points over the past few months, directly raising the financing cost of new data centers. But it does not stop there: when the system is under stress, the spread between issuance yields and the 10-year widens from 100 basis points to 150 or even 200. So every time you see rates rise, you should translate it as ‘the data centers' ability to prove out their future cash flows just got a lot harder.’ Kedrosky judges that from the financing-cost angle alone, this breaks within 6 to 12 months — the most stressed end of the market is demanding an 11% to 12% return, and that is impossible to deliver.
— Paul KedroskyData centers need to earn 7%, but nobody can make the math work
The investment-grade market currently demands a median return of about 7.2% from data centers, and the junk-grade market demands 10%. But Kedrosky stresses this is the threshold to attract creditors, not the threshold to get your capital back — that is a separate question, and he thinks it simply cannot be earned back. He coined the term EBBT — earnings before bad things — where you cut out all the spending and of course the cash flow looks great, but the premise is that hyperscalers stop burning money on frontier models, and nobody is doing that. If you only do industrial-grade inference, then it is fundamentally an energy problem: China adds roughly 3x the new energy capacity the US does each year, and is already ahead.
— Paul KedroskyAnthropic's IPO slipped from October to November, and employees are borrowing millions to exercise
What Kedrosky has heard is that Anthropic originally planned to IPO on October 1, and it is now reportedly pushed to November. His general rule: even in a good market, at least half of the companies that say they will list by year-end do not list that year. Internally there is panic about what an IPO would look like this year. More concretely, the employees: many have already exercised their options, and if the IPO does not happen they face the IRS's AMT, generating tax bills in the millions — he knows quite a few people who have borrowed millions of dollars for this. He also mentions a supplier close to the company saying Anthropic's third quarter came in below expectations, but stresses this is rumor, and consistent with the backdrop of customer concentration and token maxing fading.
— Paul KedroskyIn their own words · checked verbatim
And it's more than 60% of AI financing is now debt, is debt financed up from something like 15 to 20% a year ago. And it's now the largest piece of the investment-grade marketplace. It's now the largest piece of the high-yield marketplace. It is literally taking over global debt markets.
Paul Kedrosky4:08
So in a sense we're created this global now illusory phenomenon. that's being driven by this incredibly anomalous spending and much of which is concentrated on one specific thing, these things we, we, we call GPUs
Paul Kedrosky7:10
Back those pieces out, and by my math, the U.S. is actually in a deflationary mode, about a quarter of a percent.
Paul Kedrosky20:33
So a balance sheet is recession is a workout where heavily indebted companies, irregardless of interest rates, continue to try and get themselves less leverage, get the debt off their balance sheets.
Paul Kedrosky23:36
So every time you see rates increase, say to yourself instead, oops, it just got a lot more expensive and difficult for data centers to justify this. the future cash flows on top of which they're building these data centers.
Paul Kedrosky30:48
So if I cut out all of the things that are costing me a lot of money, and then I calculate EBBT, not EBIT, but EBBT, earnings before bad things, then my cash flow is tremendous.
Paul Kedrosky39:00
And so we're now in this moment where the companies at the center of all of this are in this kind of Minsky moment where the financialization of what they're doing is so large and so broad it's affecting everything. But at the same time, they're all very eager to dump the shares and get them out there at a credibly large scale to retail investors.
Paul Kedrosky46:10
Figures
| AI capex as a share of US GDP growth | 30% to 70% | 1:06 |
| Debt as a share of AI financing | over 60%, versus 15% to 20% a year ago | 4:08 |
| AI-related share of global goods trade | about 19% over the past three quarters | 6:10 |
| AI-related share of global trade growth | close to 55% | 7:10 |
| US inflation rate | still above 3% | 8:12 |
| Median return demanded in the investment-grade data center market | about 7.2% | 37:58 |
| Return demanded in the junk-grade data center market | about 10% | 37:58 |
| China's annual new energy capacity relative to the US | about 3x the US | 39:00 |
Glossary
- balance sheet recession
- A recessionary pattern in which companies focus on deleveraging and paying down debt regardless of the level of interest rates, leading to prolonged periods of no investment and no hiring.
- Minsky moment
- The moment when financialization decouples from underlying reality, accelerates, and then snaps.
- AMT
- A US tax law under which an employee exercising options at a low price incurs a tax liability even if the shares are not sold.
- hurdle rate
- The minimum return a project must hit; below it, investors will not put up money.
- EBBT
- Kedrosky's coinage for cash flow calculated after cutting out all the spending.
- token maxing
- Customers burning through tokens in huge volumes regardless of cost, producing pulse-like revenue growth.
How to listen
Investors and founders watching the intersection of macro and AI capex, and engineers who want to know which specific link in the chain the AI bubble cracks at first.
The small talk from 0:06 to 1:06 at the top can be skipped.