Morgan Stanley: US AI Power Gap to Reach 57 GW by 2026, Speed-to-Power Becomes New Value Frontier
nashnova research
Morgan Stanley estimates US data-center power demand will outstrip supply by 57 GW by 2026 — equivalent to six New York Cities; the ability to energize sites fast is replacing chip performance as the decisive profit lever across the AI supply chain.
Efficiency jumped sixfold — why is power still short?
AI data centers are shifting from traditional 8U servers to full-rack units packing 72 GPUs per cabinet (NVL72). Next-gen Vera Rubin and Rubin Ultra racks push power draw even higher.
Hardware efficiency is genuinely improving: token output per watt nearly sextuples between 2025 and 2028.
But Jevons' paradox — the pattern where higher efficiency drives total consumption up, not down — is at work: lower cost per unit of compute triggers explosive growth in high-load tasks like AI agents, swallowing the efficiency gains whole.
This means → chips keep getting more power-efficient, yet aggregate power demand accelerates. Supply of electricity, not silicon, is the real bottleneck.
How big is a 57 GW gap, exactly?
Morgan Stanley's figures: cumulative US data-center power demand from 2026 to 2028 reaches 97 GW. Subtract projects under construction and existing grid capacity, and the unmitigated gap is 57 GW.
Even after adding gas turbines, fuel cells, nuclear tie-ins, and repurposed mining sites, the net gap under a base case is still 33 GW — 34% of total demand.
In plain terms = 33 GW is roughly the baseload consumption of six New York Cities. Every plausible fix is already baked in, and the shortfall is still enormous.
By 2029, with the Feynman architecture coming online, the net gap widens further to 72 GW.
Why is "time-to-power" the new value frontier?
Grid interconnection approvals routinely take years. Public-grid expansion cannot keep pace with AI buildout, making "no power, no production" the binding constraint for hyperscalers.
Morgan Stanley calculates that energizing a site one year earlier creates roughly $4.50/W in incremental value — time itself is being priced.
This means → the core competitive metric is shifting from "how fast the chip computes" to "how fast the electrons arrive." Time-to-power is now the variable that determines project returns.
Who captures this windfall?
Power-shell providers (PSPs) — firms that hold ready-made land and grid-access rights and can energize data-center sites quickly — and behind-the-meter (BTM) self-generation projects are being re-rated by the market.
Many PSPs are former crypto miners. They sign 15- to 25-year leases; unlevered free-cash-flow yields on these projects run 15% to 19%.
Morgan Stanley expects CIFR, RIOT, and HUT to land major lease deals within the next six months, with West Texas emerging as a hub for BTM on-site generation.
How is profit shifting across the supply chain?
The cost mix inside a rack is changing fast: memory, high-speed interconnects, and liquid cooling are all rising as a share, while GPU's share of total rack cost keeps falling.
This reflects a broader profit migration — away from the chip layer and toward power equipment and infrastructure.
The power constraint is also redrawing the global compute map. Domestic US supply pressure is pushing hyperscalers outward; the Nordics, ASEAN, India, and the Iberian Peninsula all carry upside potential that the market has not fully priced.
Where are the risks?
Hot-section components for gas turbines face capacity constraints; equipment lead times are stretching.
Permitting delays and capex returns that fall short of projections could disrupt the industry's delivery timeline.
In plain terms = whether the gap can be meaningfully closed before 2029 is the single most important proof point for the entire supply chain — if it cannot, every compute-expansion plan downstream gets discounted.
市场有风险，内容仅供研究参考，不构成投资建议。