Morgan Stanley: 57 GW power shortfall threatens US AI buildout through 2028
Three independent forecasts and a near-unanimous House vote landed in September 2026: demand for data center power is growing far faster than the grid can be built out. The bottleneck for the AI buildout is no longer chip supply but access to electricity — and that constraint has now been both quantified and politicized.
The evidence on the table
Over roughly two weeks in September 2026, three separate forecasts landed — from Morgan Stanley, TrendForce, and Enverus — along with a House vote in which the Ratepayer Protection Act passed 417–3. Together they point to the same conclusion: the binding constraint on the AI buildout is no longer GPU availability, but how many gigawatts the grid can actually deliver. This article walks through how large the gaps are, why more efficient chips make the problem worse rather than better, how developers are bypassing the grid with gas generators, and what the bill from Congress means for who picks up the tab.
The demand: from 9 GW to nearly 80 GW in the US
According to Morgan Stanley's September analysis — known through coverage by 24/7 Wall St. — US IT power demand from data centers will rise from 9.19 gigawatts in 2025 to 78.57 GW in 2029, an increase of 755% over four years. At the same time, the bank estimates that new US data centers will need 97 GW of new power in the 2026–2028 period.
TrendForce takes a global view over a longer horizon. Its forecast, covered by Crypto Briefing, projects that global data center demand rises from 122.9 GW in 2025 to 161 GW in 2026 — a 31% increase — then to 211 GW in 2027 and 490.7 GW by 2030. By that point, available grid capacity is estimated at around 222.6 GW, leaving a global shortfall of roughly 268 GW. The US is particularly exposed: TrendForce points to a US shortfall of over 170 GW by 2030, concentrated in the grid regions PJM (the Midwest and mid-Atlantic), ERCOT (Texas), and MISO (the central US) — precisely where data centers are already clustering.
Part of the driver is visible in the composition of consumption: AI servers are expected to account for roughly 33.4% of data centers' total power demand in 2026, up from around 25% in 2025, and over 40% in 2027. The AI load is not merely displacing other data center load — it is stacking on top of it.
Why efficient chips don't solve the problem
Intuition says more efficient chips should mean lower power consumption. Morgan Stanley's analysis points to the opposite mechanism: the bank estimates that tokens generated per watt may increase almost sixfold between 2025 and 2028 — yet total power consumption keeps climbing.
This is a variant of what economists call the efficiency rebound, or Jevons paradox: when the unit cost per computation falls, the amount of computation demanded rises, and that can eat up — and exceed — the gains. This is an analysis of the mechanism behind Morgan Stanley's numbers, not something the sources explicitly conclude: efficiency gains make AI cheaper to use, which can increase usage faster than the wattage savings grow.
Hardware trends reinforce this. Morgan Stanley estimates that an Nvidia Vera Rubin rack could require around 234 kilowatts, while a future Rubin Ultra rack could reach roughly 600 kW. The shift toward rack-scale computing — in which entire rack units are delivered as a single integrated machine — was, according to the bank, part of what drove its estimate of 97 GW of new power demand in 2026–2028. The efficiency gain per compute cycle hides behind a steep rise in total installed capacity per facility.
The gaps — and why the numbers can't simply be combined
Here lies the most confusing part of the picture: Morgan Stanley estimates a US shortfall of 57 GW for 2026–2028, while TrendForce estimates a global shortfall of around 268 GW and a US shortfall of over 170 GW by 2030. These numbers must not be averaged — they are not directly comparable.
Morgan Stanley's calculation is narrower in both time and scope: the firm estimates 97 GW of need for 2026–2028 against only 19 GW of available grid capacity and 21 GW under construction, arriving at an estimated shortfall of 57 GW. According to the bank, alternative sources such as on-site gas generators, fuel cells, and nuclear-related projects could reduce the shortfall, but still leave roughly 33 GW uncovered through 2028. TrendForce projects further into the future (2030), and the coverage does not specify exactly how the firm defines demand and available capacity. The difference in time horizon and method likely explains much of the distance between the figures.
It is also important to keep the epistemic layers straight: neither Morgan Stanley's note, nor TrendForce's forecast, nor the Enverus projections are known here from primary sources — all the figures exist only through secondary coverage in the trade and financial press. That does not make them useless, but they are attributed forecasts, not verified facts.
The workaround that bypasses the grid: behind-the-meter gas
Where the grid falls short, developers are building it themselves. Enverus Intelligence Research estimates, according to coverage via Reuters and 24/7 Wall St., that 29.6 GW of gas power "behind the meter" — on-site generation connected to the facility rather than the public grid — will have been added by 2030, with data centers accounting for 88% of it.
This is a practical answer to a capacity problem: instead of waiting years for grid interconnection and transmission buildout, data centers install their own gas turbines. But it also opens new questions the available evidence does not answer: what does a significant amount of gas generation outside the shared grid mean for the AI sector's emissions? And what does it mean for grid reliability as large loads gradually detach from the common system? Those questions remain open.
The political response: the Ratepayer Protection Act
The political side of the story accelerated in parallel with the forecasts. On September 16, 2026, the House passed the bipartisan Ratepayer Protection Act by 417 votes to 3, according to CBS News and AP. House Republicans prioritized the bill, and passage required a two-thirds majority — which it achieved.
The bill requires state utility regulators to consider a principle under which power companies may charge data centers the full cost of new power generation and grid buildout needed to serve them, while preserving state regulatory authority. In other words: if a hyperscale facility requires new gas generation or new transmission lines, the cost could fall on the data center operator — not be spread across ordinary ratepayers.
The bill is not yet law. It has only passed the House, and its path through the Senate is unclear in the available source material. But it would be a mistake to dismiss it as symbolic: a 417–3 majority in a political climate marked by partisan division shows that concern over who pays for the AI buildout's power needs is broadly shared across party lines.
The political momentum is visible in polling. A survey from the AP-NORC Center for Public Affairs Research and the Energy Policy Institute at the University of Chicago shows that nearly two-thirds of Americans are very or extremely concerned about data centers' impact on energy prices, and 57% are very or extremely concerned about the impact on water supplies.
What is certain, and what is forecast
It is worth separating the layers of the source material. Verified is that the House passed the Ratepayer Protection Act 417–3 on September 16, 2026, and that the AP-NORC/EPIC poll shows widespread concern about energy prices and water. All figures about future demand, capacity, and shortfalls, by contrast, come from attributed forecasts by Morgan Stanley, TrendForce, and Enverus, known through secondary coverage.
One caveat deserves mention: 24/7 Wall St.'s own headline operates with a "$10 trillion" figure for the data center wave — but that figure is not substantiated in the source text itself and should not be treated as documented.
What to watch
Three factors will determine where this lands. First: does the Senate take up the Ratepayer Protection Act, and in what form? Second: do the differing methodologies in the shortfall estimates — Morgan Stanley's short-term US 57 GW versus TrendForce's long-term global 268 GW — converge as more data becomes available? Third: does behind-the-meter gas buildout (29.6 GW by 2030, per Enverus) grow faster than grid expansion and regulatory reform? If so, the AI industry's energy footprint moves not only off the public grid, but also out of reach of the policy Congress is now beginning to shape.
The available source material provides no answers on the emissions, pricing, or reliability consequences of this development — those are open questions that will shape the debate in the period ahead.