The chips still sell high. The loans now price the delay.

In October the local blocks stopped being proposals and became orders. A Minnesota court ordered work on Google's project halted pending environmental review. New Mexico regulators blocked the gas pipeline feeding Oracle's Project Jupiter. Governor Greg Abbott froze new grid interconnections across Texas to protect residential ratepayers ^[1]. Six months of decisions built to this. Monterey Park, California moved a citywide ban; Maine advanced a freeze on permits for facilities of 20 megawatts or more until October 2027; South Carolina's legislature took up a moratorium running to 2028 ^[2]. In July, Kathy Hochul signed New York's one-year moratorium — the first statewide ban on large centers — as opposition to facilities within three miles of homes jumped from 28% to 41%, and a survey found 53% of Americans against a data center in their own neighborhood ^[3]. By September the bans had spread through the Minnesota cities, Oregon, and at least 15 South Carolina counties, while developers in some towns resubmitted plans for a "data hall" rather than a data center to slip under the moratorium ^[4][5]. The pattern is the point, not any one town. What those towns decided has now moved inside the banks. JPMorgan, Morgan Stanley, and Bank of America formally fold community sentiment and "permitting readiness" into their credit decisions on data-center loans ^[6]. Bank of America's definition reads like a zoning hearing transcribed into a credit memo.

Readiness means all the permitting and approvals that are required, and the community support from the people who are going to live around it. — Bank of America

The banks are pricing a real gap. SynMax satellite imagery suggests roughly 40% of U.S. AI data-center projects will miss their 2026 completion dates by more than three months, and Data Center Watch counted 75 projects worth about $130 billion facing local opposition in a single quarter ^[7][6]. OpenAI says otherwise.

Our historic data center build-out is on schedule and we will accelerate from here. — OpenAI

Where the press release and the construction site disagree, the loans follow the construction site. The newest instruments share one property: they are written as if the collateral arrives late. Start with tenor, the length of the loan. New AI debt is deliberately issued at three to five years, so it expires inside what lenders see as a short hardware lifecycle, with Citadel Securities forecasting more than $500 billion more of it by 2028 ^[8]. Then the guarantor. Blue Owl couldn't place outside financing for its $4 billion Lancaster project because its tenant, CoreWeave, carries a below-investment-grade B+ rating, leaving a $500 million bridge unfunded ^[9]. The newer packages solve that by putting the chip's maker on the hook: Broadcom backstops the largest senior portions of Anthropic's $35 billion financing for Google's custom chips ^[8]. Then the price. The gap between hyperscaler — the giant cloud companies doing the building — and bank credit-default swaps has widened roughly 60 basis points, six-tenths of a percentage point, since October 2025 ^[10]. Apollo's chief economist Torsten Slok puts the repricing plainly.

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

The carrying cost has climbed at the same moment. The Fed has pushed its target range to 3.75–4.00%, five-year Treasury yields sit above 5% for the first time since 2007, and Big Tech has added $194 billion in net debt with free cash flow near zero ^[11][12]. Here is the loop: analysts at TD Securities and BMO attribute the record yields partly to the AI buildout itself — the same construction that needs financing is bidding up its own borrowing costs ^[13]. Nobody is legislating the schedule back. The president tells towns they will end up backwards and poor if they refuse a data center, while his own Treasury analysts wrote a July draft naming electricity bottlenecks among the bubble risks — an assessment the administration dismissed ^[4][14]. Elizabeth Warren wants AI companies to disclose their debt; Treasury Secretary Scott Bessent's answer is that the real risk is China pulling ahead ^[14]. The strange part is that the bulls' entire case still holds. GPU prices are spiking on a memory shortage ^[15]. Data-center vacancy sits at record lows — 0.3% in Northern Virginia — and CBRE estimates supply is at least three years from catching demand, even as its own report names limited power and community resistance as the constraint ^[16]. Equity analysts are bullish on the S&P for the fourth quarter on double-digit AI earnings ^[17]. Banks remain committed to the sector ^[6]. And roughly $4.5 trillion in private-market dry powder is still willing to fund the buildout at a price ^[18]. Every one of those statements is about price or demand. Not one buys back a day of schedule. The two markets are not contradicting each other; they are pricing different variables. The buildings fill the day they are finished. The swaps say they finish late. The market has already printed a date on that. OpenAI's roughly $1.2 trillion in take-or-pay compute commitments — take-or-pay means the buyer pays whether or not it ever uses the computing — are still in their construction phase ^[19]. 2027 and 2028 are when those contracts start actually billing, and the same window when the 2021-vintage low-rate loans need capital-structure "rightsizing," in Goldman's phrase, and must be refinanced at whatever rates then prevail ^[18]. Even the Fed's own September projections assume two rate cuts in 2027 arrive before that wall ^[12]. Two calendars are running, printed by different hands. The construction calendar is set by zoning boards, courts, and grid queues; the maturity calendar by the people who wrote the newest three-to-five-year loans. They meet somewhere in 2027 and 2028. Whether the delayed buildings start billing before the refinancing comes due is the question the debt market has already dated — it just hasn't answered it yet.