Morgan Stanley: U.S. Data Center Power Gap Reaches 32GW, Rising Order Risks for ASICs and Memory Components
nashnova research
Morgan Stanley estimates U.S. data centers face a 32GW net power shortfall from 2026 to 2028 — 34% of demand; chips can be made but not powered up, raising order-delay risk for ASICs, memory, and optical modules.
What does a 32GW power gap actually mean?
Morgan Stanley projects U.S. data centers will run 32GW short of needed power between 2026 and 2028 — roughly 34% of total demand.
This means → the AI bottleneck is shifting from "not enough chips" to "not enough electricity." Servers racked without power cannot generate compute or revenue.
Oracle's 1.3GW "Project Lighthouse" in Wisconsin is a live example — after transmission approvals were restarted, full-power delivery may slip to October 2028 at the earliest, or spring 2029 in a bear case.
Who holds up, and who gets hit first?
Morgan Stanley sees Nvidia and Broadcom 2027 guidance as largely intact — both management teams have already factored land, power, and facility constraints into forecasts, and global diversification reduces single-market exposure.
Nvidia GPUs produce more tokens per gigawatt — meaning more AI inference per unit of electricity — giving them a deployment edge when power is scarce.
In plain terms = for the same kilowatt-hour, GPUs run more workload than ASICs. When electricity is tight, customers cut the less efficient projects first — so ASICs face greater pressure.
What signal is Broadcom's guidance shift sending?
In March, Broadcom projected fiscal-2027 AI revenue would be "well above $100 billion." By September, the latest guide narrowed to $115 billion.
The absolute figure is still large, but room for further upward revision has shrunk.
This reflects expectations converging: power constraints are beginning to cap the growth ceiling.
How big is the order risk for tail-end suppliers?
Memory, optical modules, power management, and analog chips — the supporting components inside a data center — previously benefited from AI expansion and tight supply-demand.
Once end-user projects slip, customers draw down existing inventory before cutting follow-on orders — risk concentrates at the supply-chain tail.
This means → the farther a component sits from final compute output, the more exposed it is to order cancellations in a project-delay wave.
Can on-site generation and overseas expansion close the gap?
Morgan Stanley's base case adds roughly 19GW of behind-the-meter gas turbines — power plants wired directly into data centers, bypassing the grid — plus about 6GW from Bloom Energy fuel cells by 2026–2028. Nuclear and crypto-mine conversions offer further supplements, but the total still falls short of 32GW.
Offshoring helps only at the margin. Morgan Stanley has already cut U.S. share of global compute from 60% to 55%, but slow European permitting and Middle East geopolitical risk limit how much demand can relocate.
The U.S. also faces a skilled-labor shortage — the Center for Strategic and International Studies estimates the country needs more than 140,000 additional skilled workers by 2030, while the current workforce can support only about 10–20GW of new gas capacity per year.
市场有风险，内容仅供研究参考，不构成投资建议。