Skip to main content

Kansas City Fed Flags AI Debt in CRE Loans and CMBS: What It Means for CRE Investors

By Avi Hacker, J.D. · 2026-10-07

What is AI debt in CRE loans? AI debt in CRE loans is borrowing raised to build artificial intelligence data centers that has migrated into ordinary CRE credit channels: bank loan books, REITs, commercial mortgage-backed securities (CMBS), and asset-backed securities (ABS). On October 5, 2026, the Federal Reserve Bank of Kansas City put that in writing. In an Economic Bulletin titled Beyond Valuations: Long-Term Financial Risks from the AI Build-Out, economists Padma Sharma and Pierce George found that AI debt raised to finance data centers "accounts for a growing share" of all four. For the broader picture, see our guide to AI CRE finance and capital markets.

Key Takeaways

  • The Kansas City Fed states that AI data center debt now accounts for a growing share of bank CRE loans, REITs, CMBS, and ABS.
  • Special purpose vehicles formed by hyperscalers and chipmakers with private credit firms finance 90 percent of their assets with debt.
  • Those sponsors remain responsible for SPV debt through rent payments and residual value guarantees, which makes a lease the real collateral.
  • Hyperscaler and utility bonds average 16 to 17 year maturities against a 10 year market average, a duration mismatch for shorter-lived buildings.
  • The bulletin's central recommendation is disclosure, meaning this exposure is not currently visible in your lender or fund reporting.

What the Kansas City Fed Actually Published

The bulletin argues the AI build-out risk is structural, not just a stock market repricing. Sharma and George write that the build-out is financed "through diverse asset classes and novel long-term financing structures that have not yet been stress-tested by an economic downturn," and that this ecosystem "may be vulnerable to shifts in the economic outlook that are not directly related to the prospects of AI technology and its valuation."

That last clause is the part CRE investors should read twice. The warning is not that AI disappoints. It is that an ordinary recession could propagate through structures built on an assumption of continuous growth. The authors examine the full value chain: hyperscalers, semiconductor manufacturers, utilities, industrial firms, neoclouds, and data center operators. The bulletin is available from the Federal Reserve Bank of Kansas City.

One caveat. This is staff research from one regional Reserve Bank, not a supervisory rule, and the views are the authors' own.

The 90 Percent SPV: Why a Lease Is the Real Collateral

The most CRE-relevant finding is structural. The bulletin identifies special purpose vehicles as the most highly leveraged entities in the AI ecosystem. "Broad guarantees from sponsors allow SPVs to raise debt with a thin equity buffer and finance 90 percent of their assets with debt," the authors write, citing Columbia economist Stefan Van Nieuwerburgh. Neoclouds follow at 85 percent. Together these leveraged sectors carry a market capitalization of just over $3 trillion.

Then comes the sentence that should reframe any data center deal: "These firms are ultimately responsible for SPV debt through rent payments and residual value guarantees."

Translate that into underwriting language. A 90 percent loan-to-value capital stack, LTV being loan amount divided by appraised value, is not something a bank would write against a stabilized multifamily asset. It clears here only because a hyperscaler promised two things: to pay rent, and to make the lender whole on what the building is worth at the end. The credit is a tenant covenant plus a valuation backstop, not the real estate. The Kansas City Fed's contribution is showing this sits at the top of the leverage table rather than in one-off deals.

A 16 Year Bond Against a Shorter-Lived Building

The bulletin also flags a duration problem. Weighted-average maturities on AI-related investment-grade bonds run about 16 years for hyperscalers and 17 years for utilities, against a market average near 10 years. "These longer maturities suggest the AI build-out will carry enduring consequences for capital providers," the authors note.

For anyone who underwrites buildings, that gap is the interesting part. A 16 year obligation is serviced by rent from purpose-built facilities whose power density, cooling design, and chip generation may be obsolete well before the paper matures. Because financing is long-term, the authors write, these assets "will undergo revaluations in response to future technological shifts and economic cycles."

In fairness, the asset is extraordinarily tight today. CBRE's North America Data Center Trends H1 2026 report puts primary market vacancy at a record low 1.4 percent, with 80.4 percent of the 7,481 megawatts under construction already preleased. A residual value guarantee looks cheap at that vacancy, but it is priced on today's scarcity, not on year 16.

How AI Debt Reaches Your Lender, Your REIT, and Your CMBS Pool

The transmission path matters more than the headline. Because AI firms are now dominant bond issuers, the bulletin notes, "investors' purchases of corporate bonds will inevitably expose them to AI-related debt." Their investment-grade issuance rose from 2 percent of the market in 2023 to 10 percent this year, reaching $330 billion year to date, 10 times all of 2023.

Three channels deserve attention:

  • Bank CRE loan books: U.S. depository institutions were net sellers of corporate bonds in 2023 and 2024 but bought significant amounts this year. The same balance sheets write your construction and bridge debt.
  • Insurers and pension funds: Insurers, private credit firms, and pension funds have acquired long-duration AI paper through private placements, which are harder to mark and harder to see.
  • CMBS and ABS pools: Data center debt is a growing share of both, so a diversified CRE bond allocation may be less diversified than its label suggests.

This is a different claim from the volume story. We covered issuance scale in AI debt hitting $570 billion and the plumbing in AI data center debt securitization. The new point is correlation: if the same paper sits in your lender, your insurer, your CMBS pool, and your pension allocation, one repricing event touches all of them together.

What CRE Investors Should Demand Now

The bulletin's conclusion is a disclosure recommendation, not a prediction: "the opacity and complexity of emerging financial risks provide a clear case for expanded disclosures and greater transparency by financial institutions and AI firms across the entire value chain." Read as an operator, that is an admission the exposure is unmeasurable from outside. Until it is, you have to ask. Four questions to put in writing:

  • Lender concentration: What share of your lender's CRE book is data center or AI-adjacent credit, as a percentage of total CRE commitments?
  • Guarantee quality: In any data center deal you underwrite, is the credit the tenant covenant, a residual value guarantee, or the real estate? Price each differently.
  • Residual assumptions: What terminal value does the guarantee assume, and what power density and vacancy does that imply in year 16?
  • Fund-level overlap: Do your CRE debt fund, your core REIT allocation, and your insurer annuity hold the same issuers?

Note what does not change. Your DSCR math is unaffected: DSCR is NOI divided by annual debt service, and NOI excludes debt service entirely. Cap rates on stabilized assets do not move because a Reserve Bank published a bulletin. What changes is counterparty and correlation risk, the parts of the stack sitting outside the property-level model. Whether AI issuance raises your own borrowing cost is a separate question, covered in our piece on Carlyle's warning on AI financing and CRE borrowing costs.

Mapping this exposure is tedious but tractable. ChatGPT, Claude, Gemini, and Perplexity all help parse lender 10-K disclosures, REIT supplementals, and CMBS prospectuses for hyperscaler concentration, though every figure needs a human check against the source. CRE investors wanting help building that map can reach out to Avi Hacker, J.D. at The AI Consulting Network.

Frequently Asked Questions

Q: Did the Federal Reserve say AI debt threatens commercial real estate?

A: Not quite. Two Federal Reserve Bank of Kansas City economists published an Economic Bulletin on October 5, 2026 finding AI data center debt is a growing share of bank CRE loans, REITs, CMBS, and ABS. It is staff research, not a Federal Reserve policy position or supervisory action.

Q: What is a residual value guarantee in a data center deal?

A: It is a sponsor promise to cover the shortfall if the asset is worth less than an agreed amount when the financing term ends. Combined with contracted rent, it is what lets an SPV borrow against 90 percent of its assets. For a CRE investor it means the lender is underwriting the guarantor's balance sheet more than the building.

Q: Should I avoid data center exposure because of this bulletin?

A: No, and the bulletin does not suggest that. Primary market vacancy sits at a record low 1.4 percent with over 80 percent of construction preleased, according to CBRE. The useful response is to separate deals where the real estate carries the credit from deals where a guarantee carries it.

Q: How would a recession unrelated to AI hurt these structures?

A: That is the bulletin's core argument. Because sponsors finance with debt rather than equity at 85 to 90 percent ratios, a general shock can impair intermediaries holding the paper even if AI adoption keeps advancing. When firms lean on debt, the authors note, "chances increase that stresses will spill over from the firm or its sector to the broader economy."

Q: Where can I read the original report?

A: It is Beyond Valuations: Long-Term Financial Risks from the AI Build-Out by Padma Sharma and Pierce George, free on the Kansas City Fed's Economic Bulletin page. The AI Consulting Network helps CRE investors turn findings like these into an underwriting and counterparty checklist.