What is AI capex crowding out CRE borrowers? AI capex crowding out CRE borrowers is the process by which debt issuance funding artificial intelligence data centers absorbs fixed income demand, pushes benchmark yields higher, and raises the cost of capital for every other borrower, commercial real estate included. On September 15, 2026, Carlyle Group's head of global research told Bloomberg that lender behavior in AI infrastructure now echoes the mortgage market before the 2008 crisis. See our full guide to AI for CRE finance and capital markets.
Key Takeaways
- Jason Thomas of Carlyle Group said on September 15, 2026 that lenders' approach to data centers has a troubling similarity to the pre-crisis mortgage model.
- The 10-year Treasury yield briefly topped 5% on September 14, its highest level since October 2023, directly repricing CRE debt.
- Thomas also said an AI slowdown is now more likely than not, which makes CRE exposure to the AI capital cycle two-sided rather than one-directional.
- A 65% LTV loan that carried a 2.05x DSCR at 4.5% falls to roughly 1.32x at 7.0%, compressing refinance proceeds on otherwise healthy assets.
What Carlyle Actually Said on September 15
Jason Thomas, Head of Global Research and Investment Strategy at Carlyle Group, spoke with Bloomberg's Dani Burger from the sidelines of Carlyle's 2026 Global Investor Conference in Washington. He made two claims that matter for CRE. First, lenders' approach to data centers carries a troubling similarity to the model applied to mortgages before the financial crisis. Second, an AI slowdown is now more likely than not.
The analogy is about lender psychology, not asset quality. Data center credit rests on a chain of assumptions: that AI demand keeps compounding, that the compute being built finds paying tenants, and that those revenues service the debt. Concentration sharpens the risk, because a few hyperscalers drive most of the buildout. Thomas separately noted the Federal Reserve is under enormous pressure to deliver a 25 basis point hike, or a quarter of a percentage point.
Why AI Capex Crowding Out CRE Borrowers Now Shows Up in Your Rate
The crowding out argument stopped being a forecast this week and became a price. On September 14, 2026, the 10-year Treasury yield briefly touched 5.014% before easing back, its first move above 5% since October 2023, while the 30-year reached its highest level in nearly two decades. Because the 10-year is the benchmark commercial mortgage spreads are quoted over, that move passes straight through to CRE.
Several forces push the same way: heavier government borrowing, inflation pressure with Brent crude above $109 a barrel, and record corporate issuance to fund AI. Goldman Sachs expects the five largest hyperscalers to sell roughly $250 billion of bonds in 2026 and $400 billion in 2027, supply that competes for the same capital that buys CMBS and funds balance sheet lenders.
Barclays global chairman of research Ajay Rajadhyaksha framed the asymmetry bluntly, observing that the most capital intensive construction cycle in the United States is being driven by entities for whom the cost of capital is, to a first approximation, irrelevant. Hyperscalers do not stop building because rates rose 100 basis points. CRE sponsors do. We covered the early version of this dynamic in our analysis of what the bond market's AI bet means for CRE investors; the difference in September is that the pressure is now realized in quoted rates rather than projected.
The Two-Sided Trap: The Boom Raises Your Rate, a Bust Hits Your Comps
Thomas's second claim is what changes the CRE calculus. The read until now was one directional: AI borrowing crowds you out, so plan for wider spreads. If an AI slowdown is genuinely more likely than not, CRE owners are exposed on both sides of the cycle.
- If the boom continues: issuance stays heavy, benchmark yields stay elevated, and refinancing costs more for every property type.
- If the boom slows: data center land and shell comps that repriced upward lose their bid, and the construction employment that has insulated the broader economy from rate hikes softens.
Markets got a preview on Monday, September 14, when AI infrastructure stocks sold off following Anthropic CEO Dario Amodei's weekend proposal to slow the pace of frontier model development. GE Vernova fell almost 9%, Vertiv close to 8%, and Oracle almost 4%. Not everyone reads it as a cliff: Digital Realty CEO Andrew Power said slowdown pledges from Anthropic, OpenAI and xAI do not mean pencils down for AI and the real estate that supports it. But CNBC reported that new AI debt deals priced this fall are expected to clear at significantly higher rates, because fixed income investors are demanding more reward.
What This Changes in Your Underwriting Math
The practical impact is proceeds, not sentiment. Consider a stabilized multifamily asset valued at $15 million with $900,000 of net operating income, which is gross revenue minus operating expenses, excluding debt service and capital expenditures. Cap rate is NOI divided by value, so the asset trades at a 6.0% cap rate. Hold the loan constant at 65% LTV, or $9.75 million, and vary the rate on an interest only basis. DSCR is NOI divided by annual debt service:
- At 4.5%: debt service of $438,750 and a DSCR near 2.05x. Comfortable; LTV is the binding constraint.
- At 7.0%: debt service of $682,500 and a DSCR near 1.32x. Still bankable, but the cushion is gone.
- At 7.5%: debt service of $731,250 and a DSCR near 1.23x, failing a standard 1.25x test and forcing a paydown or rescue capital.
Nothing about the property changed. NOI, occupancy and market are identical. Only the cost of capital moved, for reasons originating in AI infrastructure finance rather than multifamily fundamentals. For sequencing, see our framework on ranking which loans to refinance first. CRE investors who want help wiring this into a live model can reach out to Avi Hacker, J.D. at The AI Consulting Network.
How CRE Operators Should Respond
Treat the AI capital cycle as a rate input, not a technology headline. Three moves are worth making now:
- Re-run sensitivities at a hiking Fed. Many CRE models still assume the 2025 easing path. The target range has sat at 3.50% to 3.75% since December 2025, and a September 16 hike would be the first since 2023. Our guide to AI interest rate sensitivity analysis covers the mechanics.
- Stress the bust case. If your submarket repriced on data center land demand, model that bid weakening. Avison Young reported US data center land purchases hit roughly $6 billion in H1 2026, up 79% year over year, exactly the kind of move that can reverse.
- Shorten your refinance runway. With issuance heavy and spreads widening, the option value of waiting has fallen. Price an early refinance against the carry of holding out.
Tools like ChatGPT, Claude and Gemini handle the first pass well: pulling rate assumptions out of a loan document, building a DSCR grid, or summarizing a term sheet. They are not a substitute for a lender conversation. If you are ready to make this a repeatable underwriting process rather than a one-time spreadsheet, The AI Consulting Network specializes in exactly this.
The scale of the shift is documented. Carlyle's research team asked this year whether data centers are crowding out investment in other sectors, noting that AI related borrowing accounted for roughly 30% of investment grade issuance. CBRE's H1 2026 Cap Rate Survey put a number on the pain threshold: the median respondent said the 10-year needs to reach 3.75% to lift sales volume, against roughly 4.6% at survey time. At 5%, that gap has widened, not closed.
Frequently Asked Questions
Q: Does AI data center borrowing really move commercial mortgage rates?
A: Indirectly but materially. AI related issuance competes for the same fixed income demand that funds CMBS and bank lending, which pressures spreads across the investment grade market. Most CRE debt is priced over the 10-year Treasury, and that benchmark briefly topped 5% on September 14, 2026 for the first time since October 2023.
Q: Is Carlyle saying data centers are the next subprime?
A: Not quite. Thomas compared lender psychology before the 2008 crisis, specifically the layering of assumptions about future demand and revenue durability, rather than calling data center credit equivalent to subprime. He flagged the concentration of the buildout among a handful of hyperscalers as compounding the risk.
Q: What should I change in my underwriting this quarter?
A: Rebuild rate sensitivities around a hiking Fed rather than the 2025 cutting path, stress test a scenario where data center driven land and construction demand weakens, and re-examine whether waiting to refinance still carries positive option value at current spreads.
Q: If an AI slowdown happens, is that good news for CRE borrowers?
A: It cuts both ways. Lighter AI issuance could relieve pressure on yields over time, helping refinancing. But a slowdown would also weaken data center land comps and the construction employment cushioning the economy through this rate cycle, so the net effect depends on your market and property type.