Morgan Stanley raises data center power shortfall forecast: net gap to reach 38GW by 2028

Why is the power shortfall widening further?

Morgan Stanley has raised its forecast for cumulative U.S. data center power demand in 2026–2028 from 97GW to 116GW. Before additional fast-to-deploy power solutions, the shortfall is expected to widen from 57GW to 70GW.
The increase reflects three changes to the model: including power consumption by rack components such as storage and networking that were previously excluded, raising the demand assumption for traditional non-generative AI data centers, and adjusting the power capacity required by chips based on a 75% load factor.
After accounting for solutions such as on-site generation, the bank expects a net shortfall of 43GW within the United States. After also considering the shift of some U.S. demand to Asia-Pacific, the net shortfall is 38GW, or about 33% of the 116GW in demand. Its previous net shortfall forecast was 33GW.

Morgan Stanley’s base case forecasts that 31GW of behind-the-meter power supply solutions will come online in 2026–2028, with power supplies located at data centers or close to the load. The forecast is the average of contracted projects and total potential capacity; even if equipment capacity is higher, construction labor may still constrain deployment.
The bank estimates that the existing workforce can support 10–20GW of new gas-fired generation per year, including both on-site and grid power supplies; annual additions could exceed 50GW by the end of the decade.
The model includes the transfer of approximately 6GW of U.S. data center demand to Asia-Pacific in 2026–2028, and forecasts a cumulative total of approximately 15GW by 2030. Thailand, India and Japan together have approximately 30GW of surplus generation capacity, but grid constraints limit how much can be connected in the near term.
Grid optimization and data center peak shaving could also free up capacity, but the former depends on utilities’ willingness to invest; the latter is usually short-lived and may not meet grid requirements for prolonged extreme weather.

The bank maintains its assumption of 234 kilowatts per Vera Rubin NVL72 rack: 72 GPUs at 2,300 watts each account for 165.6 kilowatts, with the remaining 68.4 kilowatts used by CPUs, networking and other rack components. Nvidia’s roadmap figure of approximately 240 kilowatts and supply-chain estimates of approximately 220–250 kilowatts provide reference points for the assumption.
The 330 kilowatts cited by Nvidia is the capacity supported by its second-generation power supply architecture, which the bank says cannot be treated directly as expected NVL72 power consumption. For Rubin Ultra, it assumes approximately 600 kilowatts, but the cancellation of the original Kyber rack design and the continued evolution of the replacement architecture make final power consumption more uncertain.
Efficiency improvements could also ease pressure on power supply: Nvidia’s power scheduling, rack power smoothing and 800V DC power distribution aim respectively to increase computing capacity within a fixed power budget, reduce peaks and cut transmission losses; these potential benefits are not yet included in the bank’s model.

Morgan Stanley expects power shortages to increase customers’ urgency to secure fast-to-deploy power solutions and powered data center sites, as well as the premiums they are willing to pay, and to drive data center growth outside the United States.
In semiconductors, the bank favors Nvidia and Broadcom on a relative basis, citing conservative guidance, visibility into chip deployment and expansion outside the United States; it also believes Nvidia generates more tokens per watt. Its model shows that although new-generation chips require higher upfront investment, lower cost per unit of compute and higher token output per watt can still support customer upgrades.
If deployments pause, downstream components such as memory, optical components and power management products could be disrupted: shortages provide price support for revenue and earnings per share in some segments, and excess ordering could also occur in a pessimistic scenario.