Blind Spots in the AI Infrastructure Selloff

Our Global Head of Thematic and Sustainability Research Stephen Byrd explains why the recent AI infrastructure selloff may reflect technical pressures, not weakening fundamentals.

Read more insights from Morgan Stanley.

----- Transcript -----

Stephen Byrd: Welcome to Thoughts on the Market. I’m Stephen Byrd, Morgan Stanley’s Global Head of Thematic and Sustainability Research.

Today: Are investors misreading the AI infrastructure selloff?

It’s Thursday, July 30th, at 10am in New York.

The recent selloff in AI infrastructure stocks has raised a familiar question: Is the buildout running ahead of real demand? The market is pulling back and we think that reflects profit-taking, crowded positioning, and forced selling by investors. This is not about weaker fundamentals. But the selloff has brought to light three key concerns, which we think the market is overplaying.

The first concern is how much enterprises are willing to pay for AI. The median enterprise employee currently generates less than $11 a month in token spending. That’s the fee paid when an AI model processes a request and generates a response.

We think there is room for that to increase. From the employer’s perspective the economics are compelling. Across workplace applications, the cost to execute the economic task would be $2-$5. And that could save an enterprise $55. That to us suggests companies are likely to spend more, not less, on AI over time.

The second debate centers on efficient models, including competitive models developed in China.  And here, policy responses both from the U.S. and China can have an impact as well. Some investors worry that better efficiency means less computing demand. But we see the opposite risk. This is a classic example of Jevons paradox: When something becomes cheaper or more efficient to use, people use more of it. In AI, lower costs can attract more users, encourage more frequent use, and make complicated applications more economical.

The scale is striking. Industry leaders estimate that compute demand could double every six months, which would amount to more than a thousand-fold increase in compute over five years. Hyperscalers could quadruple available power capacity to roughly 120 gigawatts by 2028, from about 30 gigawatts in 2025.

And that leads to the third debate – whether data centers can secure enough power to keep expanding. It’s a valid concern. In the U.S., facilities under construction and contracted grid capacity cover about 30 gigawatts. That’s less than half the 68 gigawatts of power that data centers are likely to need from 2026 through 2028. Grid connections can take five to seven years in some regions. Skilled electricians, welders, and pipefitters are in short supply. And local opposition is increasing as communities debate electricity bills, tax incentives, and who should pay for grid upgrades.

These are real obstacles, but we view them as delays rather than dead ends. Onsite generation, fuel cells, energy storage, natural gas turbines, and the conversion of existing high-power sites could close the gap, at least partially.

We believe much of the recent weakness in AI infrastructure has been driven by technical factors rather than a change in the underlying fundamentals. As AI becomes more capable and cheaper to use, demand for intelligence, compute, and power is likely to keep rising. The global market is fragmented as policy decisions in the U.S. and China shape how growth unfolds. But strong economics should support continued investment.

Thanks for listening. If you enjoy the show, please leave us a review wherever you listen and share Thoughts on the Market with a friend or colleague today.