Morgan Stanley: US Data Center Power Shortfall Reaches 34%; Leading Chip Manufacturers Temporarily Immune, Mid- and Downstream Sectors Under Pressure

According to a report from Zhitong Finance APP, Morgan Stanley has released a research report updating its estimates of the US AI data center power shortage and its industrial chain impact. The report indicates that from 2026 to 2028, the total power demand gap for US data centers will reach 57GW. After accounting for behind-the-meter (BTM) distributed generation, fuel cells, and other “plug-and-play” power solutions, the net gap remains substantial at 32GW, accounting for 34% of the total demand during the same period. Power shortages have gradually overtaken chip supply to become the core bottleneck for expanding AI computing power. However, the report emphasizes that Nvidia and Broadcom, leveraging their global presence, higher unit power-to-compute output, and clear upstream and downstream visibility, will not see their 2027 performance guidance affected by power constraints for now. Downstream segments such as ASICs, storage, optical modules, and analog devices will face bigger risks of demand volatility.

Shortfall Calculation: 34% Power Shortage to Persist Long-term, Computing Resources Concentrate Among Top Players

Morgan Stanley estimates that cumulative power demand for US data centers from 2026 to 2028 will reach 97GW, with data centers under construction covering 21GW and available grid capacity accounting for 19GW. After deducting these two, the initial power shortfall is 57GW. Factoring in the probability-weighted “plug-and-play” solutions such as BTM gas turbines, Bloom Energy fuel cells, nuclear plant site repurposing, and crypto mining site conversions, the net deficit in the baseline scenario is 32GW, or 34% of total demand; in the optimistic scenario, the gap could narrow to 15GW, while in the pessimistic scenario, it could widen to 42GW.

Year by year, the power gap is set to expand: the net deficit is about 9GW in 2026, around 15GW in 2027, growing to 32GW in 2028, and further rising to 68GW in 2029. The report notes that large-scale deployment of next-generation high-power GPU racks (such as NVL72) is the main driver behind surging power demand. On the supply side, the landscape is also highly concentrated. From 2026 to 2028, hyperscale cloud providers and emerging cloud companies (Amazon, Google, Microsoft, Meta, etc.) will contribute 60%-95% of global data center capacity additions. Leading players, by virtue of stronger creditworthiness, revenue predictability, and purchasing scale, hold a dominant advantage in securing power resources, while smaller cloud providers, international companies, and low-efficiency chip manufacturers face the risk of being squeezed out.

(In 2027, hyperscale and emerging cloud providers will add 32GW of computing capacity, and 39GW in 2028; among them, Google will rank first in new operational capacity additions.)

Diverging Industry Impact: Leading Chipmakers Safe, Downstream and Tail-end Under Pressure

The report argues that power shortages have a clearly tiered impact on the semiconductor supply chain. Nvidia and Broadcom together control about 90% of the AI XPU market, with a combined 12-month AI revenue guidance of approximately $800 billion, and this guidance has not been substantially threatened by power shortages so far.

Their core support stems from four factors: management’s performance outlook has fully considered risks of site, power, and infrastructure (LPS) shortages; they have strong visibility into chip deployment locations across the chain; aggressive expansion of global computing nodes outside the US; and Nvidia’s unit GW computing output is significantly higher than its peers, providing a marked comparative advantage in a power-constrained environment.

Relatively, ASIC chips have lower token output per unit of power, making their market share susceptible to displacement by GPUs in power-constrained scenarios. At the end of the supply chain, low-value segments such as storage, optical modules, power management, and analog devices are most susceptible to the bullwhip effect—if AI expansion slows, customers will first delay or cancel orders for these components, increasing revenue and profit volatility risks for related manufacturers.

Limited Relief Pathways, Overseas Migration Unlikely to Fill Domestic Shortfall

Regarding possible relief measures for the deficit, the report believes that overseas AI computing expansion cannot make up for US domestic shortages.

Although US companies are accelerating the deployment of computing capacity in Australia, Asia, and Europe, limitations such as power constraints and lengthy approval cycles in Europe, geopolitical risks in the Middle East, and the expectation that Asia can only absorb 14GW of US supercomputing overflow demand by 2030 mean that the overall scale is insufficient to offset the US domestic gap. As a result, the report lowers the projected US share of global computing power from 60% to 55%. Behind-the-meter (BTM) distributed generation is currently the primary incremental power source. In the baseline scenario, BTM gas turbines and engines will contribute 19GW of power from 2026-2028, with a maximum of 49GW in the most optimistic scenario. However, factors such as skilled labor shortages, engineering complexity, and local approval procedures render the realization of these capacities highly uncertain. Additionally, fuel cells and nuclear site repurposing offer some supplementary supply, but it will be difficult to fundamentally resolve the long-term shortfall.

Overall, power shortages are reshaping competitive dynamics in the AI compute and semiconductor supply chains. The trend towards resource concentration among leading companies will further strengthen, giving an edge to businesses with energy efficiency advantages, global footprints, and access to power resources, while mid- and downstream niche sectors should be wary of demand volatility risks.