Morgan Stanley raises its estimate of the U.S. AI power gap: names power plant targets that "shorten the energization cycle," with NVIDIA Corporation (NVDA.US) becoming the top semiconductor pick under the "power shortage bottleneck."

In a thematic research report published on October 26, Morgan Stanley raised its estimate of the net U.S. data center power gap for 202628 to 38GW from 33GW. At the same time, Morgan Stanley noted that amid the power gap, ASICs may lose share relative to GPUs. The trigger for this round of updates is that investors are trying to reconcile capex expectations of US$1.5 trillion in 2027 and US$1.6 trillion in 2028 with power availability. Morgan Stanley stated bluntly that about one-third of the power gap is not market consensusmany investors believe companies will "creatively find ways around the bottleneck," and cite still-strong supply chain orders as evidence. The bank's stance has not wavered: it remains confident that the power gap is real and is a key constraint on construction. It acknowledges that there are three levers that can narrow the gapbehind-the-meter power, spillover to Asia, and grid optimization and data center peak shavingbut each faces its own constraints; at the same time, it defended its Vera Rubin rack power assumptions. How is the 38GW calculated? The gap in the research report is not a single number, but three layers. The first layer is the "pre-solution gap": 202628 U.S. data center power demand of 116GW (previously 97GW), after deducting 25GW of in-construction front-of-meter and behind-the-meter computer rooms and 21GW of available front-of-meter grid capacity, leaves a gap of 70GW (previously 57GW). The second layer is the "U.S. domestic net gap": after including five categories of "shorten the energization cycle" solutions, the report gives low/medium/high scenariosadditional behind-the-meter turbines and engines (6/22/38GW), Bloom Energy fuel cells (5/6/8GW), co-locating data centers at operating nuclear power plants (1/3/8GW), repurposing existing energized computer room sites (11/19/27GW), and spillover of U.S. demand to Asia-Pacific (5/6/8GW). After probability weighting, the net gap is 60/43/23GW, representing 52%/38%/20% of demand. The third layer is the "net gap including global solutions": after also including Asia-Pacific spillover, it is 56/38/17GW, representing 48%/33%/14% of demand. The report marks the median scenario with "we believe this is the most likely outcome"that is, 43GW (38%) in the U.S. domestic case, or 38GW (33%) including global solutions. This is exactly where this upward revision lands. Why did demand jump from 97GW to 116GW? Three model changes pushed demand higher: first, the utility power expected to be needed to power all purchased chips was raised, adding energy consumption from storage, networking, and other rack-level componentsthese were previously not included; second, total power demand from historical non-generative AI data centers was raised; third, correctly incorporating a 75% load factor in the power methodology increased the total power required for chips to run at full load. The report also emphasized the need to clarify three power measurestotal power (the total amount delivered by the grid), critical IT power (the portion delivered to chips after PUE conversion), and utilized power (what chips actually "consume," including the load factor)and to convert correctly among the detailed line items in order to derive total demand. The upward revision was partly offset by more front-of-meter grid capacity (rising from 15GW to 20GW in 202628) and behind-the-meter solutions, but the net result is still a widening gap. There are five ways to close the gap, but each has a ceiling It is worth noting that the report assigns a probability of success to each lever: additional behind-the-meter turbines and engines 90%, Bloom Energy fuel cells 90%, co-location at operating nuclear plants 70%, repurposing existing energized computer room sites 85%, and spillover to Asia-Pacific 90%. The probabilities are not low, but no single lever can close the 70GW gap on its own. Even in the median scenario, the five combined still leave a 43GW hole. The report's subtext is clear: the gap can be narrowed, but it cannot be eliminated. The ceiling for behind-the-meter power is labor, not manufacturing capacity Morgan Stanley's base case forecasts 31GW of behind-the-meter solutions coming online in 202628, taking the average of "signed contracts" and "total potential capacity"; the report acknowledges that if total turbine capacity, etc., is used, the gap could narrow further. But the bottleneck then shifts to labor. Its utilities team believes the existing labor base can support only about 1020GW per year of new gas additions (including both behind-the-meter and grid-side); however, by the end of the decade, this number could rise to more than 50GW per year. In other words, behind-the-meter power is not "build as much as you want," but is constrained by how many people can install it. Asia has power, but may not be able to connect it The second challenge points to spillover. The report estimates 6GW of spillover in 202628; the Asia team, through conversations with companies, expects that by 2030 about 15GW of U.S. hyperscale demand may spill over to Asia, with another about 3GW from new fabs in ASEAN, India, and South Korea. By region: Malaysia 4.0GW, other ASEAN 4.0GW, India and others 3.4GW, Australia 2.0GW, Japan 1.5GW. Thailand, India, and Japan have the strongest data center commitments, and Thailand also benefits from demand from Chinese hyperscalers. These markets together have about 30GW of surplus power capacity, from ramping gas, coal, and nuclear capacitybut the report also points out that grid constraints limit near-term serviceability. Having power does not mean it can be connected. Grid optimization and peak shaving: theoretically workable, constrained in reality The third challenge is the most technical, and Morgan Stanley's rebuttal is also the most blunt. Grid-enhancing technologies (software and hardware that make the grid "smarter") are indeed a lower-capex path to raising the capacity of existing lines, but the key constraint is whether utilities are willing to invest. Data center peak shaving (committing to use less grid power during peak demand) could in theory relieve grid pressure, but the report believes utilities will not increase data center interconnection capacity because of itbecause they manage the grid under the most extreme conditions, and the duration of demand response is far shorter than the extreme events they plan for. The report gives a very concrete example: during a winter storm with rolling blackouts, data centers are unlikely to run diesel generators for six consecutive days, especially given state and local emission limits on pollutants and particulate matter. In addition, there are almost no regulatory actions in the U.S. that could achieve this result. The key question: how many kilowatts does Vera Rubin actually use? In September, Morgan Stanley raised Vera Rubin rack power to 234kW from 149kW and Rubin Ultra to 600kW from 415kW, and received considerable feedback; this time it responds specifically. Its calculation logic is: Vera Rubin NVL72 assumes a single-GPU thermal design power of 2,300W, with 72 GPUs totaling 165.6kW, plus another 68.4kW for CPU, networking, and other rack components; for Rubin Ultra, it uses the 600kW figure proposed by NVIDIA Corporation at GTC 2025. Four reference points support this number: NVIDIA Corporation's own roadmap (reiterated at last week's OCP Summit) puts NVL72 at about 240kW; the 330kW also cited by NVIDIA Corporation is the power capacity supported by its 110kW power shelves, not the expected power consumption of NVL72; the MGX 1.2 architecture supports up to 250kW; on the supply chain side, Schneider Electric gives 227246kW, Vertiv about 240250kW, TrendForce about 225kW, and Pegatron specifies 228kW for its NVL72 Max-P. The report also mentions three types of engineering improvements worth noting, though the model does not currently reflect their potential benefits: NVIDIA Corporation's MaxLPS (on-demand power allocation to GPUs), rack-level intelligent power smoothing (using energy storage and GPU-level control to flatten power spikes), and 800VDC power distribution. For Rubin Ultra, the report acknowledges that because the original Kyber rack design was canceled and replaced by a new architecture, the final configuration is still evolving, with greater uncertainty. Who benefits? The report believes the gap will increase the urgency and premium for two types of names: first, U.S. "shorten the energization cycle" solutionsbehind-the-meter suppliers Bloom Energy (BE.US), Solaris (SEI.US), and energized computer room suppliers TeraWulf (WULF.US), Hut 8 (HUT.US), Cipher Mining (CIFR.US), and Riot Platforms (RIOT.US); second, data center growth outside the U.S. In semiconductors, it remains most bullish on NVIDIA Corporation (NVDA.US) and Broadcom Inc. (AVGO.US), citing conservative guidance, visibility on chip deployment, geographic expansion outside the U.S., and NVIDIA Corporation's higher token output per watt. If construction pauses, the disruption is most likely to appear at the end of the "bullwhip effect"that is, where shortage-driven pricing supports revenue and profit, such as memory, optical modules, and other downstream components, as well as power management and analog devices that could be over-ordered in a bear case. The report also warns that in a power-constrained environment, ASICs may lose share relative to GPUs because of their lower token output per watt and weaker cross-customer substitutability. The cost-side logic is: although total data center cost rises with each generation, unit cost falls fasterdata center cost per TFLOPs has fallen from about US$14.30 for Hopper to about US$0.88 for Rubin Ultra. On this basis, the report argues that even with the gap, companies still have incentives to adopt solutions that shorten the energization cycle. It should be noted that the report's disclosure page shows: Morgan Stanley states that it has business dealings with the companies it covers and may have conflicts of interest; as of September 30, 2026, it held more than 1% of the common stock of multiple companies including NVIDIA Corporation, Broadcom Inc., Bloom Energy, GE Vernova, TeraWulf, and Solaris; in the past 12 months, it managed or co-managed public offerings for companies including NVIDIA Corporation, Broadcom Inc., Bloom Energy, GE Vernova, TeraWulf, Solaris, Equinix, and NextEra, and received investment banking service compensation from multiple companies; in addition, several signing analysts and research personnel personally hold shares in companies mentioned in this report.