Morgan Stanley: U.S. Data Center Power Gap Equals Six New York Cities
nashnova research
Morgan Stanley estimates the US data-center power shortfall will reach 30–40 GW by 2028 — equivalent to six New York Cities' base load. Without mitigation, it swells to ten. AI's capacity buildout is pushing power supply toward a systemic bottleneck.
How large is the gap?
Under Morgan Stanley's base case, the US data-center power supply-demand gap hits 30–40 GW by 2028. New York City's base load runs about 5.5–6 GW — the gap equals six New York Cities.
Without any mitigation, the shortfall widens to the equivalent of ten New York Cities.
This means → this is not a single-plant expansion problem; it is a structural deficit across the entire grid.
Why does a more efficient chip burn more power?
The key driver is the power leap in Nvidia's next-gen architecture: Vera Rubin and Rubin Ultra server IT power draw is projected to grow fivefold from 2025 to 2028.
Switching from the Hopper series to Rubin Ultra cuts the all-in cost per processor operation by 94% and lifts performance per watt roughly sixfold.
In plain terms = when the cost per unit of compute drops that sharply, operators don't save electricity — they pack more GPUs into the same rack space. This is the Jevons paradox (higher efficiency → higher usage → total consumption rises).
CoreWeave disclosed that a typical Vera Rubin NVL72 rack generates heat equivalent to 10 to 25 ovens running at once.
What happens after 2029?
Morgan Stanley projects that once Nvidia's next-next-gen architecture Feynman ships in 2029, US power demand will jump another 40% above the 2028 baseline.
The power gap would then widen to roughly 107 GW — about 18 New York Cities.
This means → even if the 2028 gap is partly closed, the next chip generation will immediately rip it open again.
What are the fixes — and are they credible?
The report lists several "fast-connect" paths: Bloom Energy fuel cells could add 5 GW by 2028; behind-the-meter direct connections at nuclear plants could contribute another 2 GW.
Even so, the base case still shows a 17 GW cumulative shortfall in 2028, dependent on unannounced gas-turbine and off-grid generator projects to close.
By 2029, the portion relying on solutions that do not yet exist commercially grows to 24 GW.
In plain terms = existing fixes cover roughly half the hole; the other half is a bet on projects that haven't broken ground.
Who is placing early bets — and where does the investment thesis point?
Morgan Stanley sees the investment thread running through any company with a viable energy-production or sales solution, including "power-shell providers" — firms previously focused on Bitcoin mining that are pivoting to supply data-center electricity.
Elon Musk spent roughly $1 billion of personal funds in July to acquire off-grid gas and diesel generator supplier APR Energy; SpaceX recently broke ground on its own gas-turbine blade foundry in Texas.
This reflects a broader signal: whether power supply can keep pace with chip shipments is now viewed as the core variable for validating AI capex — not a bonus, but a prerequisite.
市场有风险，内容仅供研究参考，不构成投资建议。