When AI Needs Less: Efficiency Risk in the Data Center Debt Boom
- Ben Rockmuller
- Feb 18
- 3 min read
The AI boom is no longer just an equity story, where tech companies fund out of revenue and retained earnings. It is now a debt story as well. Moody’s estimates that data centers will require at least $3 trillion of investment through 2030, spanning servers, buildings, and power infrastructure¹. A growing slice of that is coming through the investment-grade market: banks expect hundreds of billions of new AI-linked bond issuance from the major hyperscalers in 2026.
The complication is this: what happens if the AI world eventually needs less hardware than today’s plans assume? In tech, progress shows up in two places: the code gets more efficient, and the chips get faster. New model architectures and training techniques can deliver the same results with far fewer computations, while each new generation of hardware tends to do more work per watt and per dollar. Put simply, smarter algorithms can run on better machines and may not require all the capacity we are racing to build today. We saw a version of this with the recent DeepSeek scare, when software and services stocks shed roughly $800 billion of market value in six trading days on fears that more efficient AI would cannibalize existing business models².
For bondholders, the concern is obsolescence risk. Many of these projects are being financed with 7-, 10- or even 30-year debt on the assumption of a long, stable economic life. If future AI models can do more with much less compute, some facilities may prove too large or too power-hungry. Leverage lurks behind assets whose true economic life could be much shorter than anticipated.
Even if the first stress shows up in private credit or project loans, the theme can bleed into investment-grade credit through more highly levered balance sheets, data center landlords and utility contracts tied to a small set of hyperscale tenants, and a general repricing of “AI infrastructure” risk across the capital structure. Oracle’s recent $25 billion bond deal, part of a broader plan to raise $50 billion (half debt and half equity), drew more than $120 billion of orders, so on the surface it looks like mega-cap tech spreads can stay contained³. But in our view, this is not a comforting bellwether: for Oracle, the equity piece is doing a lot of work to preserve the firm’s rating, and tech spreads broadly are heading wider (chart). Moreover, we are concerned that additional leverage is being created in off-balance-sheet structures like special-purpose vehicles and private securitizations where a new entity builds the data center, takes on the debt, and signs guaranteed contracts with an anchor tenant.
In Beyond the Headline Yield, we showed that nominal yields overstate what investors earn. Here, we think it is important to look past the AI headline and focus on what is being promised to bondholders over time. In 2026, our bias is to be cautious on issuers whose balance sheets (or contingent obligations) are increasingly tied to long-dated data center spending, and to place more weight on well-capitalized, diversified sectors such as large banks, brokers and consumer staples companies, whose fundamentals are less directly exposed to whether the next generation of AI needs more hardware… or a lot less.
Comments