By: Russ Kamp, CEO, Ryan ALM, Inc.
I’ve been spending a lot of time recently thinking about the incredible amount of capital being thrown at Artificial Intelligence (AI). There is no question that AI has the potential to dramatically change how we live and work. But does that mean that every dollar being invested in AI infrastructure is going to generate an acceptable return? I’m not so sure. As interest rates continue to rise, I think investors need to start asking a different question. Instead of asking, “How big can AI become?”, perhaps we should be asking, “How much AI capacity can actually be built economically?”
AI investment is staggering. Microsoft, Amazon, Alphabet/Google, Meta, and Oracle (the Hyperscalers) are at the center of this massive infrastructure buildout. Until recently, these companies generated so much cash that they could fund most of their capital spending internally. That situation is changing rapidly. AI-related capital expenditures are consuming an enormous percentage of their operating cash flow, and the hyperscalers are increasingly turning to the bond market and other forms of financing to keep the spending machine going. Four of the five major hyperscalers have issued significant amounts of bonds in 2026, with Microsoft being the notable exception. Collectively, hyperscaler borrowing has already reached roughly $200–$220 billion this year. Wow!
Why should we care? Because the cost of money matters and this massive investment could profoundly impact equities, bonds, real estate, private equity, private credit, etc. DB Pension plans need to take notice! When Treasury yields were 1%-2%, financing massive data centers and other AI infrastructure was relatively inexpensive. Today, Treasury yields have climbed above 5%, and the 30-year Treasury bond is closing in on 5.7%, while some long-term hyperscaler debt is being issued at 6% or more. Oracle has recently issued long-dated debt carrying coupons approaching 8%. This reality changes the economics dramatically. It isn’t enough for a $10 billion or $20 billion AI project to generate revenue. It needs to generate a return sufficient to compensate investors for the cost of the capital, operating expenses, electricity, depreciation, technological obsolescence, and the risk associated with the project. What is that return likely to be and where is that return going to come from?
Capital isn’t the only potential constraint. AI requires enormous amounts of electricity, generation capacity, transmission, transformers, land, cooling, water, semiconductors, construction, and, of course, data centers. A large AI campus can require hundreds of megawatts of electricity. High-density data centers can cost roughly $14-$16 million per megawatt, meaning that a 500 MW facility could cost approximately $7.5 billion to construct before considering the broader power infrastructure necessary to support it. We keep hearing about seemingly unlimited demand for AI. Fine! But there certainly isn’t unlimited electricity, grid capacity, construction capability, or CAPITAL. Why does the investment community seem to assume otherwise?
Furthermore, the AI trade may be creating its own headwind. Think about this for a minute. Massive AI capital expenditures consumed free cash flow from most of the hyperscalers. Declining or exhausted free cash flow creates a need for external financing. Greater borrowing produces more corporate bond supply. More bond supply can contribute to higher yields and wider credit spreads. Higher financing costs increase the hurdle rate on the next AI project. Eventually, some projects simply won’t make economic sense. That seems like the beginning of a vicious cycle to me.
As mentioned previously, the implications extend well beyond technology stocks. Equity investors need to determine whether these enormous capital expenditures are actually producing an acceptable return on invested capital. Bond investors are being asked to absorb hundreds of billions of dollars of new AI-related debt and need to be compensated appropriately. Real estate investors financing data centers must compete against a >5% risk-free Treasury yield while dealing with higher construction and financing costs.
I’m certainly not suggesting that the AI boom is about to end. But I do believe that the AI investment thesis may be entering a very different phase. The first phase was about AI models, semiconductors, hyperscalers, and data centers. The next phase may increasingly be about the scarce resources necessary to support all of that growth, beginning with capital and including electricity, generation, transmission, transformers, cooling, powered land and water. The winners may ultimately be those controlling the scarce resources rather than simply those spending the most money.
Markets have an interesting habit of believing that trends can continue indefinitely. From my 45-years in the investment industry, I’ve come to appreciate that they don’t. There is always a natural capacity to every investment. AI will prove to be no different. At today’s cost of capital, the important question isn’t how much AI infrastructure companies want to build. It is how much they can afford to build while still generating an acceptable return. Are today’s investors and pension plans adequately considering that distinction? I’m not convinced that they are.
I have an idea. While you wait for the AI thesis to play out, buy time (extend the investing horizon) by creating a cash flow matching (CFM) portfolio that will secure the monthly benefits and expenses for some time – say 10-years. This will enable that AI thesis to perhaps generate the desired return while it grows unencumbered.
