In this edition of Beyond Buildings: The F Factor, Logan Freeman of Midwest CRE Advisors argues that the CoreWeave credit selloff was a financing event, not a technology verdict. Drawing on a recent Kansas City Fed bulletin, he explains why tenant credit, documented utility progress and lender underwriting now decide which data center sites and powered shells hold value, and what landowners, developers, investors and operators should ask as credit tightens.
Last week the bond market did something the stock market has not done in this AI cycle. It sent a bill.
CoreWeave, the largest of the so-called neoclouds, issued bonds earlier this year with coupons of roughly 9.6% to 9.75%. During the week of October 12, those bonds traded down hard enough that buyers were earning close to 15%. Five-year credit default swaps, which are insurance against a default, went above 1,000 basis points. Bloomberg reads that as about a 60% implied chance of default within five years. The stock lost about a quarter of its value in five trading days, and the Nasdaq fell 4.5% for the week.
Before anyone writes the obituary, be clear on what did not happen. Nothing defaulted. CoreWeave still pays the same fixed coupons it agreed to. Nvidia has agreed to buy unsold CoreWeave capacity through April 2032. Third-quarter earnings come in November, and that report will tell us more than any headline.
What changed was the price the market charges to lend into this story. For anyone who owns land near a substation, builds powered shells or invests in data center real estate, that is the part that matters.
Ray Dalio said it plainly on LinkedIn last month:
"The danger is that people fail to distinguish the miracle from the investment. A technology can be truly transformative, but if the stocks are priced too high or bought on debt, you get a crash."
He went further on October 7 at the Forbes Global CEO Conference in Singapore. He called AI a "classic bubble" because of the concentration and the amount of debt being taken on to fund it while rates rise, and he said, "We're in the part of the cycle that is before that but approaching that." A year earlier, in October 2025, he had told CNBC that bubbles don't really pop "until they are popped by tightness of monetary policy."
That is the framework I keep coming back to. AI is real. The compute demand is real. Bubbles don't come from the technology. They come from the money stacked on top of it, and from what happens when that money gets more expensive or wants out.
CoreWeave borrowed roughly $50 billion to buy GPUs and fund data centers that take years to pay back, and it did that while rates were rising. When the cost of capital moves against a business that depends on capital markets to grow, the market reprices the debt first. That is what we saw last week. In my opinion, it is a financing event, not a technology verdict.
The most useful lesson for real estate people is inside CoreWeave's own capital stack.
In March, CoreWeave closed an $8.5 billion facility, its first investment-grade GPU-backed financing, rated A3. It was non-recourse and secured by GPUs, a contract with a single investment-grade customer, and the related data center leases. The pricing was about 5.9% fixed, or SOFR plus 225 basis points.
In August, a $2.6 billion facility priced at SOFR plus 5.5%.
Same company, same year, same kind of collateral, and very different pricing. One clear difference was what sat behind the loan. When the loan is tied to an investment-grade customer under a long contract, lenders treat it like infrastructure. When it depends mostly on the sponsor's own credit, they treat it like a speculative bet.
Goldman Sachs' chief credit strategist, Amanda Lynam, wrote last week that as AI issuance has grown, "construction risk, coverage of principal or refinancing risk, and claim to tenant have become increasingly important," along with power availability, permitting and insurance. That list is basically a data center underwriting checklist.
Tenant credit now drives financeability. For years, people asked whether there was enough power for a data center. The question now is who signs the lease and whether a lender will finance against their signature. A powered shell leased to an investment-grade hyperscaler and the same building leased to a thinly capitalized GPU renter are not the same asset, even if the slab, the switchgear and the substation are identical.
Leases and GPU rentals are different risks. A data center lease is a real estate obligation. A GPU rental or compute contract is a service agreement with equipment that can become obsolete quickly. When a deal stacks one on top of the other, the landlord's real credit is often the end customer's, not the operator's. Underwrite the whole chain. The Kansas City Fed's October 5 bulletin found that neoclouds run debt-to-asset ratios of about 85%, and that the special purpose vehicles in this ecosystem finance about 90% of their assets with debt. Thin equity is fine until the refinance comes due.
Utility timelines don't care about capital cycles. Speculative money moves in months. Substations, interconnection studies and transmission upgrades move in years. Capital that bought land on the assumption of fast power is the capital most exposed when financing tightens. My opinion: sites with documented utility progress, meaning a study, a load letter or a queue position, will hold their value better than sites selling on proximity to a line on a map.
Lenders will slow down and sort. Expect more questions, more diligence on tenant credit and less tolerance for speculative shells. I don't have clean cap rate data to cite yet, so I won't make any up. But the direction of credit pricing is clear from the bond market, and real estate usually follows credit.
The Kansas City Fed bulletin makes the railroad comparison directly. The railway build-out gave rise to the American corporate bond market starting in the 1830s, but the financing strains didn't show up until 1873. The railroads got built anyway. Many of the people who financed them lost money.
Telecom ran the same play. Global Crossing spent a reported $13.7 billion building a global fiber network and filed for bankruptcy in January 2002 with $12.39 billion of debt. By 2002, a TeleGeography figure cited by the Wall Street Journal put the share of installed U.S. fiber actually in use at 2.7%. The OECD found that telecom accounted for 56.4% of the $163 billion in corporate bond defaults worldwide that year. The internet was real. The debt load wasn't sustainable.
Office is our own industry's recent version. A building is only worth what it can be financed for at today's rates, against today's tenants.
These are questions, not investment advice. Here is what I would be asking right now:
CoreWeave may come through this fine. Nvidia is in its corner and the backlog is real. But the market just reminded everyone that the technology and the investment are two different things. In data center real estate, the investment is the lease, the power and the credit behind both.
Where are you seeing lenders push back hardest right now: tenant credit, power timelines or construction risk?
Copyright 2026, Midwest CRE Advisors
