The party's not over, but someone just asked to see the bill.
The Summary
- Investors have poured $300 billion into AI debt across every credit instrument, but now they're getting picky about who gets the next round of funding.
- "Choosiness" is emerging as lenders start asking uncomfortable questions about when all this compute spending turns into actual profit.
- The shift signals a maturation from "fund everything AI" to "show me the unit economics."
The Signal
Three hundred billion dollars is a staggering number for what amounts to a two-year infrastructure bet. That's more than the GDP of Finland, deployed into data centers, GPU clusters, and the energy grids to power them. The money flooded in from everywhere: investment-grade bonds, leveraged loans, private credit, asset-backed securities. If it had a yield and an AI pitch deck attached, it got funded.
Now the music is slowing. Winnie Cisar, global head of strategy at CreditSights, says investors are showing "choosiness" for the first time since the AI debt boom began. This isn't a pullback yet, it's a recalibration. The difference matters.
"After a $300 billion binge, choosiness is just another word for 'prove it.'"
Here's what changed:
- The first wave of AI infrastructure debt went to anyone with Nvidia on the purchase order
- Wave two funded model training companies burning cash on compute
- Wave three is asking: when does revenue exceed the electric bill?
The fatigue isn't about AI skepticism. It's about debt fundamentals reasserting themselves. Credit investors want to know the path from training runs to cash flow, from agent demos to paying customers, from "we're building AGI" to "here's our debt service coverage ratio." These are reasonable questions that weren't getting asked six months ago.
The timing matters because we're entering the part of the AI buildout where capital intensity meets revenue reality. Data centers are online. Models are trained. Agents are shipping. Now comes the hard part: do customers pay enough to cover the interest payments? The $300 billion already deployed isn't going anywhere, those bonds are issued, those loans are on the books. But the *next* $300 billion will require better answers.
The Implication
If you're raising AI debt in the next twelve months, your pitch deck better have a revenue model that doesn't rely on the greater fool theory of compute resale. Investors are still bullish on AI infrastructure, but they're no longer writing blank checks. This is healthy. It separates companies building real agent platforms from companies renting GPUs to other companies renting GPUs.
Watch for spread widening on AI debt from companies with pure capex stories versus those showing usage-based revenue traction. The companies that can demonstrate they're selling *outputs* (agents that do work) rather than *inputs* (raw compute) will get the next round of funding at better rates.