Finance

AI Build-Out Hits a Shortage of Cheap Money

Marcus SterlingPublished 3w ago4 min readBased on 9 sources
Reading level
AI Build-Out Hits a Shortage of Cheap Money
Photo by Kevin Ache on Unsplash

Lenders and investors are getting choosier about financing the AI build-out and demanding stricter terms as the build-out enters a new phase, according to The Wall Street Journal, which published its report on October 31, 2026. The constraint is no longer only chips, power and land. It is the price and conditions of the capital that pays for them.

The volume of projects needing funding is large and growing fast. Reuters reported, citing John McWilliams, that U.S. data center capacity under construction rose 48% between December 2025 and June 2026, measured in gigawatts. Each gigawatt under construction needs a financing stack of equity, bank debt, bonds and sometimes private credit, and each layer is now being assembled under tighter scrutiny. The Wall Street Journal has described the build-out as on track to become the biggest economic bet in U.S. history.

Pricing is the most visible signal. Bloomberg reported that at the start of 2026, the highest-rated data center bonds offered yields below 5%. Yield is the annual return a bond pays against its price, so a rising yield means investors are demanding more compensation for the same paper. The starting point is useful.

Lenders are also pricing in the physical side of the build. Reuters Breakingviews reported that lenders are charging more and demanding tougher terms to protect themselves against supply-chain bottlenecks. Those bottlenecks are the kind that delay equipment and push completion dates, and a late project is a project that services its debt later. Senior bankers told Reuters they scrutinize community concerns when assessing U.S. data center project loans, which adds a permitting and reputational dimension to credit work that lenders have not historically had to model at this scale.

Underwriting has become harder for a more basic reason. Bloomberg reported that investors are struggling to underwrite data center deals because requirements for cooling and computing density are evolving rapidly. Underwriting is the process of judging whether a borrower can repay and whether the collateral will hold its value. If the specification of a building changes while it is being financed, the asset that secures the loan is harder to value on the day the loan is made and in the years after. That is a different problem from a weak borrower, and it is not solved by a higher spread alone.

Some of the exposure may not yet be visible. S&P Global estimates that the world's biggest tech firms likely have more AI-linked liabilities than currently known, as reported by Reuters. Off-balance-sheet structures, leases and special-purpose vehicles are the usual places where such obligations sit, and analysts who rely on reported debt figures may be working from an incomplete picture.

The leveraged finance market, where loans and bonds go to borrowers carrying high debt relative to earnings, is also part of the story. Bloomberg reported that Morgan Stanley expects about $20 billion of AI-related deals in leveraged finance markets in 2026.

For expert readers, the open question is whether stricter terms slow the build or mainly reprice it. If lenders demand higher spreads, tighter covenants and more equity, projects with thin sponsor support or uncertain tenants will struggle to close, and the marginal developer will feel it first. If the largest sponsors can still raise capital at workable terms, the adjustment lands on smaller players and on the cost of the marginal gigawatt. The reporting so far points to repricing and selectivity more than a halt, but the distinction matters for anyone holding the debt, underwriting it or depending on the capacity it finances.

For ordinary savers and borrowers, the connection runs through the bonds and loans held in pensions, insurance portfolios and mutual funds. Those holdings are where a change in the cost of AI infrastructure credit would eventually surface.