70GW "Power Gap"! U.S. Power Grid, Citizens, and AI Giants—One Side Has to "Make Concessions," Massive Capital Expenditure "Cannot Be Realized"

The explosive growth of artificial intelligence is now fiercely colliding with the physical limits of the U.S. power grid. The resulting massive power shortage is posing a substantial threat to the infrastructure expansion plans of global tech giants. The severe mismatch between supply and demand means that the colossal AI capital expenditure anticipated by the market faces the grim prospect of being “unrealizable,” as the power grid, the public, and tech giants are now engaged in a high-stakes battle.

According to the latest estimates from Morgan Stanley analyst Stephen Byrd, U.S. data center developers face a power shortfall as high as 70GW by 2028. Even after exhausting all potential solutions—including nuclear energy, fuel cells, and shifting some demand to Asia—the net power gap still stands at a staggering 43GW. This figure has more than doubled in just a year, highlighting the hard constraints physical infrastructure imposes on the AI industry’s development.

This supply-demand imbalance is rapidly reshaping market pricing and capital flows. On one hand, equipment vendors capable of delivering “time to power” solutions are commanding extraordinary premiums, with prices of core equipment like gas turbines having soared by twofold. On the other hand, as power costs rise and local opposition intensifies, credit markets are becoming pickier about AI-related debt. Investors are increasingly concerned that some large-scale data center construction plans may eventually have to be delayed or canceled.

Amid soaring power costs, public opposition to new data centers in the U.S. is mounting, and multiple state governments have begun to freeze project approvals. The market faces a harsh reality: either tech giants will be forced to scale back their grand capacity plans, or they will pay an unprecedented price in terms of delays and costs to secure power access.

Despite the latest data from the U.S. Energy Information Administration (EIA) showing that the U.S. grid achieved its strongest capacity growth in a decade in September, this masks a deeper structural crisis. The types of new capacity, their geographic location, and their grid connection schedule are entirely mismatched with the relentless, round-the-clock demand of data centers. The entire industry has fallen into a web of bottlenecks formed by equipment shortages, labor gaps, and policy resistance.

70GW Shortfall: Power Remains the Most Critical Constraint

Morgan Stanley’s AlphaWise survey clearly exposes the core of this crisis.

The survey shows that “speed of power access” is cited as the number one bottleneck by global data center operators, accounting for 27%. In the U.S., this proportion is as high as 45%, far exceeding any other country. “Availability of powered land” comes second, with the U.S. also leading globally at 26%.

Morgan Stanley’s calculation framework is as follows: U.S. data center developers will cumulatively need 116GW of power by 2028. Data centers under construction (including independent power projects) cover 25GW, and available grid access covers 21GW, creating a gross 70GW shortfall. Adding in “quick power” solutions—such as suitable nuclear sites (2GW), announced turbines and engines (20GW), Bloom fuel cells (5GW), and shifting 5GW demand to Asia—the net shortfall still reaches 38GW, about 33% of the total required power.

Morgan Stanley’s U.S. data center team, expressing figures using utility metrics, states: 2028 demand at 180GW vs. 137GW of supply—producing a 43GW deficit, nearly three times greater than the current 15GW gap.

Demand Surge and "Ineffective" Supply: Grid Structure Mismatch

Wall Street investment banks are now dramatically raising expectations for power demand.

Bernstein analyst Sunaina Ocalan has raised her CAGR forecast for U.S. power demand before 2030 from 3% to 4%.

The firm predicts that from 2025 to 2030, the U.S. will see approximately 1,000 terawatt-hours (TWh) of new power demand, with more than half coming from data centers.

Goldman Sachs’s model also shows that by 2030, U.S. data center power demand will jump from 39GW in 2025 to 108GW.

However, structural mismatches on the supply side are casting doubt over this demand growth. Goldman’s data shows the U.S. grid added a net 5.6GW of generating capacity in September, but mainly from solar (3.0GW) and battery storage (1.6GW). For data centers requiring 24/7 operation, solar’s capacity factor is only about 25%—incapable of supplying stable baseload power.

Natural gas power capacity, which could meet this demand, is seriously lagging. In September, there was zero new capacity added from combined-cycle gas plants, the main source of baseload power. Goldman points out the concentrated delivery period for gas power projects won’t arrive until 2028–2030, which misses Morgan Stanley’s forecasted shortfall peak. This timing mismatch means that in the coming years, the grid simply will not be able to provide enough suitable power to support AI growth.

The Twin Shackles of Supply Chains and Labor

Even with abundant funding, the physical world’s supply chain and labor shortages present nearly insurmountable obstacles.

On the equipment side, Morgan Stanley notes that GE Vernova’s gas turbine orders are now backlogged until 2029, with prices having soared from $1,000 per kilowatt in the early 2020s to $3,000. Delivery times for large power transformers are as long as 3–4 years, while switchgear requires 1–2 years. Across-the-board equipment cost increases are pushing up data center construction expenses, which are expected to rise from $30 billion per GW in 2022/23 to $50–60 billion by the end of the decade.

More challenging than equipment is human resources. Goldman Sachs estimates that by 2030, the U.S. electric power sector will require more than 500,000 new workers. Even under the most optimistic apprenticeship assumptions, there will still be a shortage of 78,000 workers in transmission and distribution by 2030.

Bernstein analyst Chad Dillard’s model, based on existing electricians, plumbers, and HVAC technicians’ construction capacity, shows that the labor shortage could become the most rigid constraint, capping IT capacity deployment at 142GW in 2030—well below optimistic market expectations.

Tech Giants’ Self-Rescue and the Rise of Behind-the-Meter Generation

Faced with grid slowness, tech giants are now adopting self-rescue strategies by bypassing the traditional grid. Goldman Sachs predicts that by 2030, behind-the-meter (BTM, i.e., power not flowing through the public grid) generation will power 25% of data centers. Morgan Stanley is even more aggressive, forecasting U.S. BTM demand at 47GW by 2030.

Morgan Stanley’s baseline scenario assumes around 20GW of new BTM gas turbines and engines—with a range of 6 to 38GW—added before 2028, plus about 6GW of under-construction installations and 6GW of Bloom fuel cells. Its U.S. data center team tracked 25–30GW of manufacturer orders for 2026 to 2028, and expects developers to add more than 15GW of BTM capacity per year starting 2028. Yet even under the most optimistic “quick power” bundle scenario, there would still be a 17GW net shortfall (38GW under the baseline).

Nuclear energy has also become a highly sought-after solution. Tech giants like Google and Amazon have recently signed frequent agreements with nuclear operators such as Constellation, attempting to secure power by restarting or upgrading existing reactors. However, these are no quick fixes. Morgan Stanley notes that delivery for new large nuclear plants is around 15 years, and commercial deployment of small modular reactors (SMRs) won’t arrive until the mid-2030s—far too late to fill the 2028 gap.

In this process, tech giants are almost insensitive to power prices. Morgan Stanley’s analysis shows that even if power prices rise 17% (from $120/MWh to $140/MWh), the return on equity (ROE) for a 100MW AI data center would only drop by about 60 basis points, still staying above 30%. Such indifference to prices shifts pressure onto ordinary taxpayers and the grid system.

Public Backlash and Policy Resistance: Who Pays the Bill?

While tech giants are willing to pay high electricity prices, ordinary citizens and voters are not. Morgan Stanley estimates that global electricity prices have risen about 20% this year. In the latest four capacity auctions by PJM (the largest U.S. regional grid operator), data centers accounted for 46% of the costs.

Soaring power bills have triggered strong political backlash. Gallup’s latest poll shows that 70% of Americans oppose building data centers in their local communities, and this opposition crosses party lines. State governments have begun to act: Texas ordered a halt to new data center construction in September and launched an audit of all projects in ERCOT (Texas grid) interconnection queues; Pennsylvania scrapped its fast-track approval process; New York suspended environmental permits for data centers above 50MW.

Policy resistance is directly causing a halt in capital expenditure. Morgan Stanley’s data shows that about $156 billion worth of data center projects were canceled or delayed in 2023, with another $130 billion in the first quarter of 2024. As AI-related debt issuance hit $471 billion this year, credit markets are becoming increasingly cautious.

If physical power cannot be delivered, massive capital expenditure plans risk being shelved. With data centers expected to consume one-seventh of U.S. electricity by 2030, among the grid, the public, and AI giants, one party is bound to be forced to compromise.