A pullback that looks contradictory until you look closer
Reports citing Simply Wall St and Reuters say Morgan Stanley is reducing its exposure to data center construction loans, even though the bank joined Bank of America, Citi, and Blackstone on October 26 in a 63 billion dollar financing package for Anthropic's AI computing capacity, built on Broadcom technology, with deliveries expected to begin in 2027. On its face that looks like one hand pulling back while the other leans in.
It is not actually contradictory once you separate the two decisions. The pattern is banks shifting AI infrastructure credit risk toward private credit funds and institutional investors rather than exiting the sector outright, while still participating directly in the highest quality, best collateralized deals like the Anthropic package. Morgan Stanley is not leaving AI financing as a business line. It is becoming more selective about which AI financing it holds on its own balance sheet versus what it originates and syndicates out to other investors willing to carry the risk.
The numbers behind the selectivity
The repricing is stark. Top rated data center bond yields were below 5 percent at the start of 2026 and were near 7 percent by October, and Morgan Stanley has cited around 8 percent as the level where financing becomes much harder to place. Goldman Sachs projects global hyperscaler debt issuance will reach 420 billion dollars in 2027, up roughly 68 percent from about 250 billion dollars in 2026, a growth rate that is outpacing investor appetite for the credit risk behind it.
Credit spreads on 10 year hyperscaler bonds widened about 50 basis points in 2026, versus roughly 20 basis points for high quality industrial borrowers over the same stretch, a gap that simply did not exist a year earlier in either market. The five year US Treasury yield also rose more than a full percentage point in 2026, raising the baseline cost of every dollar borrowed for data center construction regardless of the borrower's credit quality, before any sector-specific risk premium even gets added on top.
Why now, and why it is not really about AI demand
The timing lines up with a broader risk-off mood across tech markets generally. The Nasdaq 100 lost about 2.5 trillion dollars in market value in a single session during the November selloff, and that kind of volatility makes lenders more conservative across the board, not only toward AI-specific credits they were already watching closely. Morgan Stanley's own research frames the concern narrowly rather than as a call that AI demand itself is collapsing or that the buildout has been a mistake.
The risk it flags is a timing and underwriting problem, not a demand problem. Projects financed during the most optimistic stretch of the AI buildout could struggle to earn their expected returns if construction costs keep climbing or customer commitments shift before the underlying debt matures years from now. A lender manages that kind of risk by tightening terms and selecting deals more carefully case by case, which is a meaningfully different response than exiting a market altogether.
Power availability is the wildcard behind the financing math
Underneath the credit numbers sits a physical constraint lenders are now underwriting against directly, whether or not they say so publicly. US data center power demand could grow from roughly 60 gigawatts today to 150 gigawatts by 2028, but grid and construction constraints could leave actual delivered capacity 30 to 40 percent below that ambition on current trajectories. A loan underwritten against a facility's projected revenue only performs as modeled if the power actually arrives on the schedule the borrower promised investors and lenders alike.
That gap between announced capacity and deliverable capacity is precisely the kind of uncertainty that pushes a careful lender to tighten terms rather than simply charge a higher rate and move on to the next deal. Morgan Stanley reportedly trimming its loan book reads as a direct response to that specific uncertainty about delivery timing, rather than a broader bet against AI infrastructure as a category worth financing at all, which is a meaningfully different signal for borrowers to read.
Neoclouds are the pressure point to watch
The squeeze is already visible among the neoclouds that depend most heavily on external financing. Nebius obtained a 3.5 billion dollar senior secured loan from Citi in October after dropping a planned high yield bond, a pivot that only makes sense if the bond market terms it was seeing had deteriorated. Early 2026 also saw about 22 billion dollars in asset backed and commercial mortgage backed securities issued against data center assets, a sign that financing is fragmenting across more structures as straightforward corporate debt gets harder to place at acceptable cost.
Banks are using tools like significant risk transfer to keep client relationships intact while shifting part of the potential losses onto outside investors, which lets them stay active in the market without carrying the full exposure on their own books. That structure works until the point where it does not, and it is worth remembering that preliminary risk transfer discussions first reported as early as December 2025 should not be mistaken for completed, battle-tested arrangements that have actually been through a stress event.
What this means for your capacity contracts
If your organization has signed, or is negotiating, a multi-year capacity commitment with a neocloud or a smaller infrastructure provider, the counterparty's financing structure is now a legitimate line item in your risk review, not just its technical roadmap or its chip supply agreements. Ask directly what share of the facility's construction is bank-financed versus private credit, and what covenants or triggers could force a renegotiation of your contract if that underlying financing gets repriced before your term ends.
This is not a reason to avoid neocloud capacity altogether, which remains genuinely useful for workloads that do not need hyperscaler-grade redundancy or geographic diversity. It is a reason to treat the financing stack behind a capacity deal the way you already treat a vendor's own financial health, as something your procurement and legal teams review carefully before signature, not something you discover only after a lender's pullback has already become your delivery delay.



