Not one price for every dollar
I keep coming back to this: if AI becomes a very attractive place to put money, what happens to everything else asking for that money?
But first, a correction to my own framing. There isn’t one universal hurdle rate. A speculative compute project and a Treasury aren’t interchangeable alternatives. Opportunity cost means comparing investments of equivalent risk; the hurdle also depends on how a project is funded. Damodaran’s starting point is useful here.
Bonds offer liquidity, income, and liability matching. Credit spreads compensate for default risk, among other things. Neither has to “beat compute” at being compute.
My takeaway: AI may change some funding decisions. That is a smaller, more useful claim than saying it sets the price of all money.
The competition is still worth watching
A big investment wave can compete for savings, lenders’ balance sheets, and attention. It can also generate income and attract new capital. Crowding out isn’t something I get to assume just because the spending is large.
The Fed’s July 2026 report describes strong AI-related investment and accommodative financing for large businesses alongside more restrictive conditions for small firms. It also notes AI-related debt issuance.
That coexistence is interesting. It isn’t evidence that AI caused the small-firm squeeze. Monetary policy, borrower quality, and banks’ own constraints could explain it.
I’d want to see a connection between the additional funding demand and the terms offered to other borrowers. Otherwise I’m joining two observations with a good story.
The borrower at the edge
If there is a squeeze, I’d look at access before averages.
A smaller credit line. A refinancing that doesn’t happen. A business deciding not to hire because the loan no longer makes sense. Those seem more informative than a single headline bond yield.
Small businesses, weaker corporate borrowers, and dollar borrowers abroad are possible places to look. Not a ranking of who gets hurt first. Different lenders, currencies, collateral, and repayment schedules matter.
The distinction I want to keep: capital can remain plentiful for one borrower while becoming scarce for another. The next question is why, not whether the average looks fine.
Cheaper tokens don’t settle it
Cheaper inference could unlock enough new uses to increase total compute demand. It could also reduce the hardware needed to serve existing demand. Calling the first outcome “Jevons” doesn’t tell me which wins.
More usage doesn’t automatically mean better returns, either. Competition can pass the savings to customers.
And the downside isn’t just some discounted metal with resale value. A compute investment can lose its capital. Obsolescence, low utilization, and power commitments matter; debt and leases add obligations even when revenue disappoints.
Microsoft’s FY2026 Q4 discussion distinguishes cash spending from finance leases and identifies CPUs and GPUs as short-lived assets. I need to follow those obligations, not just the headline capex number.
The loop I’m trying to test
Here’s the part that feels counterintuitive.
Suppose AI investment makes financing harder for some other borrowers. Suppose that weakness then spills into hiring and spending. If growth and inflation soften enough, risk-free rates might fall.
Could cheaper financing then support another round of AI investment, even while other businesses remain under pressure?
Possibly. But every “suppose” is doing work. Lower policy rates need not mean lower long-term yields. And lower Treasury yields need not mean cheaper financing if credit spreads widen or equity investors demand more compensation for risk.
Lisa Cook’s discussion of AI and interest rates captures the competing forces: investment demand could raise the neutral rate now, while later effects could run the other way. The direction isn’t settled.
- AI investment draws more funding If new capital doesn’t fully meet that demand
- Some other borrowers lose access If the squeeze materially affects hiring and spending
- Broader demand weakens If inflation and market expectations allow
- Risk-free rates fall If risk premiums don’t offset the fall
- AI financing becomes cheaper
Possible return to the start: more AI investment, but only if expected customer demand and project returns still justify it.
Failure path: weaker demand also hurts AI revenue → expected returns fall → projects are delayed or cancelled. The loop stops.
Plenty of ways this doesn’t happen
The obvious one: customers cut AI budgets too. A labor-saving tool might appeal during a slowdown, but adoption still needs money and a credible payback. Cost-cutting demand isn’t recession-proof.
Or inflation stays sticky. Or long yields rise. Or electricity and construction capacity constrain the buildout before funding does.
Funding structure matters as well. Spending from operating cash flow, issuing debt, and signing a long lease aren’t the same thing. Internally funded projects still have an opportunity cost, but they don’t need to refinance every dollar in a credit market.
And the simpler explanation might win: government borrowing or ordinary credit deterioration could matter much more than AI capex.
What I’d watch, and where I land
- Credit availability for smaller and weaker borrowers, not only average spreads.
- Cash generation against spending, debt, and lease commitments at the builders.
- Actual customer revenue and utilization against the assumptions behind new capacity.
- Risk-free yields and borrower-specific financing costs separately.
The open question: can I distinguish this mechanism from an ordinary uneven credit cycle? Watching two things move together won’t be enough.
Where I land for now: an AI boom and financial stress elsewhere can coexist. A slowdown might even help finance parts of the boom. That possibility is worth tracking; it isn’t yet a reason to assume the loop is real, permanent, or tradable.