Morgan Stanley urges buying AI infrastructure dip after 7% slide

Morgan Stanley says the rout in AI infrastructure stocks — from memory chip makers Micron Technology and Intel to construction giant Caterpillar — is a technical correction, not a broken thesis, urging investors to buy the dip after shares fell an average of 7 percent in July.

"This point in time represents an unusually attractive buying opportunity," analysts at Morgan Stanley wrote in a July 28 note, arguing AI infrastructure will become an "intelligence superhighway" delivering significant net benefits to economies worldwide.

The bank's 120-page report, "Playing the AI Infrastructure Dip," attributes the pullback to position crowding, margin deleveraging, and momentum reversal rather than a breakdown in fundamentals. It estimates US data center power demand of about 68 gigawatts for 2026-2028, leaving a potential 38GW gap after accounting for 15GW under construction and 15GW of contracted grid capacity. Grid interconnection queues in some regions now stretch five to seven years.

The report identifies "Powered Shell" providers — companies with land and power access that can be converted to data centers — as the most mispriced segment. Names including TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, and Galaxy Digital trade at $2-4 per watt of enterprise value, versus 20-25x for mature operators like Equinix and Digital Realty. Morgan Stanley assigns the group a 15x EV/Watt target.

Token economics and the Jevons Paradox

The report pushes back on two market fears. First, that companies capping employee AI token usage points to slowing demand. Morgan Stanley counters that a single AI call saves about $55 in labor costs while an agent-completed enterprise task costs $2-5, a return above 10x. Tools with that economics are a competitive necessity, not a budget line item, the bank argues.

Second, that China's Kimi K3, trained at lower cost, undermines the rationale for hyperscaler capital spending. Morgan Stanley invokes the Jevons Paradox — the 19th-century observation that efficiency gains raise total consumption. Cheaper compute expands use cases, and the bank cites a Google executive's estimate that compute may need to double every six months, a 1,000-fold increase over five years. Nvidia's AI chip sales, growing at a roughly 140 percent compound annual rate from 2025-2028, would still cover less than a tenth of that single customer's projected demand.

GPU generations also protect margins. Morgan Stanley's Intelligence Factory model puts net margins on token sales at about 58 percent for Blackwell-based data centers, rising to roughly 80 percent for Rubin and 90 percent for Feynman. That cost-curve shift lets hyperscalers cut token prices by about 75 percent without sacrificing profitability, the bank said.

The 38GW power gap and who fills it

The report concedes physical constraints are real, grouping them as "3P" — people, power, and politics. Skilled trades are in short supply, grid queues stretch five to seven years, and state and federal politics are turning against data center construction. The House is reviewing the Ratepayer Protection Act, which would codify a national data center electricity surcharge after Amazon, Google, Meta, Microsoft, Oracle, and xAI signed a White House pledge in March.

Morgan Stanley frames these as speed bumps, not structural barriers. It sees bitcoin miners converting existing grid capacity and land into data centers, adding 10-19GW, while gas turbines contribute 15-20GW and fuel cells 5-8GW. Even with those "time-to-power" solutions, the bank's mid-case still leaves a roughly 1GW net gap, widening to 11GW in a bear case.

The scarcity is not capital but powered physical space. Morgan Stanley's 15x EV/Watt target on Powered Shell providers implies substantial upside from current $2-4 levels, though the group faces execution and regulatory risk. For hyperscalers, the power bottleneck — not chip supply — is now the binding constraint on data center growth.

This article is for informational purposes only and does not constitute investment advice.