Morgan Stanley has released a research report updating its calculations on the U.S. AI data center power shortage and its impact on the industry supply chain.

The report shows that the total U.S. data center power demand gap for 2026-2028 reaches 57GW. After deducting "plug-and-play" power solutions such as behind-the-meter (BTM) generation and fuel cells, the net gap still stands at 32GW, accounting for 34% of total demand over the same period.

Power shortages have gradually replaced chip supply as the core bottleneck for AI computing expansion. However, the report emphasizes that NVIDIA and Broadcom, leveraging their global footprint, higher computing output per unit of power, and clear upstream and downstream visibility, will see their 2027 earnings guidance remain unaffected by power constraints, while midstream and downstream segments such as ASICs, memory, optical modules, and analog devices will face greater demand volatility risks.

Morgan Stanley estimates that cumulative U.S. data center power demand for 2026-2028 will reach 97GW, of which data centers under construction cover 21GW and available grid capacity covers 19GW, leaving an initial power gap of 57GW after deducting these two items. After factoring in probability-weighted "plug-and-play" solutions including BTM gas turbines, Bloom Energy fuel cells, nuclear plant site conversions, and crypto mining site conversions, the net gap under the base case is 32GW, representing 34% of total demand; under the optimistic scenario, the gap could narrow to 15GW, while under the pessimistic scenario, it could widen to 42GW.

On an annual basis, the power gap shows a trend of widening year by year: the net gap is approximately 9GW in 2026, about 15GW in 2027, expanding to 32GW in 2028, and further climbing to 68GW in 2029. The report notes that the large-scale deployment of next-generation high-power GPU racks, such as NVL72, is the core driver behind the surge in power demand.

The supply side simultaneously exhibits a highly concentrated landscape. Among global new data center capacity from 2026 to 2028, hyperscale cloud providers and neocloud providers (such as Amazon, Google, Microsoft, and Meta) will contribute 60%-95% of the incremental capacity. Leading players, with stronger credit profiles, revenue certainty, and procurement scale, hold an absolute advantage in the competition for power resources, while small and mid-sized cloud providers, international players, and low-efficiency chip manufacturers face the risk of being squeezed out.

The report believes that the impact of power shortages on the semiconductor supply chain is significantly tiered. NVIDIA and Broadcom together hold approximately 90% of the AI XPU market, and their combined 12-month AI revenue guidance of approximately $800 billion is not currently under substantial threat from power shortages. The core support comes from four aspects: management's earnings guidance has fully accounted for site, power, and infrastructure (LPS) shortage risks; high visibility into chip deployment locations across the entire chain; active global computing node deployment outside the United States; and NVIDIA's significantly higher computing output per gigawatt compared to peers, giving it a clear comparative advantage in a power-constrained environment.

In contrast, ASIC chips have lower token output per unit of power, and their market share faces the risk of being squeezed by GPUs in a power-constrained environment. At the end of the supply chain, low-value segments such as memory, optical modules, power management, and analog devices are most vulnerable to the bullwhip effect — if computing deployment pace slows, customers will prioritize delaying or canceling orders for these components, exposing related manufacturers to higher revenue and earnings volatility risks.

Regarding mitigation pathways that the market is focused on, the report argues that overseas computing expansion cannot fill the domestic U.S. gap. Although U.S. companies are accelerating computing deployments in Australia, Asia, Europe, and other regions, Europe is constrained by power limitations and approval timelines, the Middle East carries geopolitical risks, and Asia is expected to absorb only 14GW of U.S. supercomputing demand spillover by 2030, with the overall scale insufficient to offset the domestic U.S. gap. The report accordingly downgrades its expectation for the U.S. share of global computing from 60% to 55%.

Behind-the-meter (BTM) generation is currently the most significant source of incremental power. Under the base case, BTM gas turbines and engines can contribute 19GW of power from 2026 to 2028, with an optimistic scenario reaching up to 49GW. However, constrained by factors such as skilled labor shortages, engineering complexity, and local approval processes, the actual achievable scale carries significant uncertainty. Additionally, fuel cells and nuclear plant site conversions can provide some supplementary power, but cannot fundamentally resolve the long-term gap issue.

Overall, power shortages are reshaping the competitive landscape of the AI computing and semiconductor supply chains. The trend of resources concentrating toward leading manufacturers will further strengthen, and companies with energy efficiency advantages, global deployment capabilities, and power resource acquisition capabilities will benefit more, while midstream and downstream niche segments need to guard against demand volatility risks.