Electricity remains the hardest ceiling on AI expansion.
A September 21 report from the trading desk indicates that Morgan Stanley's latest research reveals the U.S. data center power deficit is far more severe than market expectations, driven by the accelerating adoption of rack-scale AI architectures. The firm projects cumulative U.S. data center electricity demand of 97 gigawatts (GW) from 2026 to 2028. After subtracting 21 GW from data centers under construction and 19 GW of available grid capacity, the power shortfall reaches 57 GW before any solutions are applied—roughly 50% larger than the prior estimate of 38 GW.
Even after incorporating "rapid power access" options such as natural gas turbines, Bloom Energy fuel cells, and co-location projects at nuclear plants, the median net deficit remains 33 GW, representing 34% of U.S. data center electricity demand during the period. The bank vividly describes the scale: "The electricity we lack equals the base consumption of six New York Cities (about 5 to 6 GW each)."
The shift to rack-scale architecture is the direct driver of the widened gap
The substantial upward revision of the power deficit stems primarily from a fundamental change in how AI chips are deployed.
The industry is moving from server architectures based on 8 GPUs to rack-scale designs with 72 GPUs per rack (NVL72). This transition is not just a change in chip count but a reconfiguration of the entire system's power logic—each rack must integrate high-speed interconnects, memory, liquid cooling, and power management, pushing power consumption far beyond the simple sum of chips.
Morgan Stanley analysts have accordingly raised their power assumptions sharply:
- Vera Rubin: per-rack IT power increased from 149 kW to 234 kW
- Rubin Ultra: raised from 415 kW to 600 kW
- Feynman: lifted to 1,100 kW
These changes boost U.S. power demand forecasts for 2027 and 2028 by 38% and 69%, respectively, versus the previous model. Morgan Stanley defines 2027 as the "inflection point" for rack architecture—when the industry shifts from optimizing individual chips to optimizing entire compute systems.
How much of the gap can rapid power access fill
To narrow the shortfall, Morgan Stanley analysts incorporate several "rapid power access" solutions, including behind-the-meter generation, Bloom Energy fuel cells, and building data centers adjacent to operating nuclear plants.
In the base case, the probability-weighted supply contribution from these options is estimated at 24 GW, reducing the 2026–2028 deficit from 57 GW to 33 GW.
Breaking it down:
- New behind-the-meter gas turbines and reciprocating engines: 19 GW
- Bloom Energy fuel cells: 6 GW
- Data centers co-located near operating nuclear plants: 3 GW
- Some crypto mining site conversions are counted within data centers under construction or new behind-the-meter generation, and are not treated as an additional reduction
The report's low, base, and high scenarios show net deficits of 42 GW, 33 GW, and 0 GW for 2026–2028; the 33 GW figure is deemed the "most likely" outcome, equal to 34% of U.S. data center deployment demand in the period.
Analysts also note that the higher net deficit partly reflects a change in model calibration: the report no longer double-counts Powered Shell Providers (projects with grid connection ready and shell structures available) as a "rapid power access" reduction, since these are now included in data centers under construction, grid connections, or behind-the-meter generation.
The economic value of rapid power access: one year early, what is it worth
According to Morgan Stanley's model, securing electricity one year earlier creates value equivalent to roughly $4.50 per watt, which is 5.9 times the annual electricity cost.
The reasoning is straightforward: billions of dollars in hardware sitting without power is idle capital; delaying power means deferring compute revenue and missing the competitive window for current-generation chips.
The report states: "The urgency among AI players to secure power far exceeds market perception. Any 'rapid power access' supplier that can deliver electricity to data center developers in 2027–2028 will be able to sign highly attractive contracts."
Former bitcoin mining firms that have transformed into Power Shell Providers (PSPs) are direct beneficiaries of this logic. These companies build data center "power shells" at a cost of $10–12 per watt, sign fixed-price contracts spanning 15–25 years, and achieve unlevered free cash flow yields of 15%–19%. Morgan Stanley calculates that at a 15x EV/EBITDA valuation, net value creation per gigawatt reaches $1.55 billion—roughly 15 times that of renewable energy PPA models.
Year-to-date 2026, the number of PSP transactions has already surpassed the full-year 2025 total.
Beyond grid connections: on-site power and pre-powered sites gain attention
Amid the 57 GW raw deficit and the 33 GW base-case net shortfall, analysts categorize solutions that shorten the time-to-power as critical mitigation paths.
The report notes that on-site generation can bypass the "often lengthy and uncertain" grid interconnection process; fuel cells can provide site power before grid hookup is complete; and pre-powered sites with existing grid access can be converted into high-performance computing data centers.
Analysts expect more behind-the-meter generation deals in the coming months, particularly in West Texas. Meanwhile, site providers with grid connection resources may also monetize capacity by building and leasing out high-performance computing data centers.
Analysts also anticipate that, due to U.S. power shortages, hyperscale cloud providers and other AI players may increase overseas data center expansion, with markets such as Nordics, Iberian Peninsula, Australia, ASEAN, and India potentially absorbing more incremental demand.
The report concludes that as high-density racks accelerate deployment after 2027, the primary data center concern will shift from "whether there is enough floor space" to "whether each rack can be supplied with sufficient power and cooling."