AI infrastructure's power crisis is spawning an energy supply revolution. Goldman Sachs believes that, constrained by severely lagging grid expansion, the data center "behind-the-meter power plant" model is moving from a marginal option to an industry mainstream, and is expected to bear 25% of global data center power demand by 2030—a share that was nearly zero in 2025.

On September 27, Goldman Sachs' Carbonomics team, in a 50-page report, sharply raised its global data center behind-the-meter (BTM) generation capacity forecast from 40GW to 67GW, and estimated that gas turbines, reciprocating engines and fuel cells combined will supply 28% of U.S. data center power demand and 25% of global demand.

At the same time, the report pointed out that the gas turbine market is already under severe strain—GE Vernova's backlog plus reserved capacity has reached 116GW, 2030 capacity is fully sold out, and more than 50% of 2031 capacity has also been sold. This supply bottleneck is pushing data center operators toward alternative solutions such as fuel cells, reciprocating engines and even industrial boilers.

Over the longer term, Goldman Sachs believes small modular nuclear reactors (SMRs) are the only scalable permanent solution after 2030, but due to construction cycles, their large-scale deployment is unlikely to be realized before 2030.

Demand Explosion: Data Center Power Demand Nearly Triples in Five Years

The AI computing arms race is pushing global power grids to their limits. Goldman Sachs' U.S. technology team and GS SUSTAIN team forecast that global data center installed capacity will rise from 101GW in 2025 to 217GW in 2030, with the United States alone reaching 108GW.

Measured by electricity demand, Goldman Sachs expects global data center electricity consumption in 2030 to increase 170% compared with 2025, a sharp upward revision from its previous forecast of 117%, with more than 60% of the incremental growth coming from the United States.

This trend will push Goldman Sachs' forecast for the compound annual growth rate of total U.S. electricity demand to 3.5%—a number that would have been regarded as fantasy at any utility investor day five years ago.

Grid Under Pressure: New Transmission Lines Plunge, Interconnection Wait Approaches Five Years

The pace of grid expansion is far from keeping up with demand growth. The Goldman Sachs report shows that the average annual mileage of new U.S. high-voltage transmission lines has plunged from 1,700 miles in 2010-2014 to 350 miles in 2020-2023, with only 55 to 125 miles added between 2023 and 2024.

Meanwhile, the median waiting time from submitting an interconnection application to formal commercial operation has approached five years. Goldman Sachs believes data from power equipment manufacturer INNIO is even more severe: grid connection time has extended from about two years historically to more than seven years today.

It is precisely this reality that has prompted hyperscale data center operators to begin signing 15-year BTM power contracts. When waiting for grid connection means a multi-billion-dollar campus cannot power up until the next decade, "temporary" self-generated power naturally evolves into a permanent solution.

Political pressure also cannot be ignored. The report cited data showing that this summer, 142 anti-data-center protests broke out across 42 U.S. states. Every GW of self-generated power is one GW that will not be spread across ordinary residents' electricity bills, and the political cost of the BTM model is continuing to decline.

Fuel Cells Win: Speed Premium Overwhelms Cost Disadvantage

In Goldman Sachs' comprehensive scoring system, fuel cells rank first with 76.6 points, ahead of aeroderivative gas turbines (68.2 points), reciprocating engines (67.0 points) and heavy-duty gas turbines/combined cycle units (59.6 points)—even though their levelized cost of electricity (LCOE) is the highest among all options.

The cost comparison is clear at a glance: for a 500MW data center, Goldman Sachs estimates that the LCOE of reciprocating engines is about $80/MWh, combined cycle gas turbines (CCGT) about $81/MWh, open cycle gas turbines (OCGT) about $91/MWh, and fuel cells as high as $117/MWh—even with a 30% investment tax credit (ITC), this only falls to $90/MWh.

However, the core competitiveness of fuel cells lies in "time." In Goldman Sachs' scoring system, "time to power" carries a weight of 20%, and "availability" carries a weight of 15%, while LCOE and initial capital expenditure together account for only 25%. Solid oxide fuel cell (SOFC) manufacturers quote only 6 to 12 months from order to power supply; reciprocating engines require 1.5 to 2.5 years; heavy-duty gas turbines require 5 to 7 years, far above the normal market's 2 to 3 years.

Bloom Energy's remarks on its second-quarter earnings call precisely captured this logic: customers now measure return on investment by "total power cost per unit of computing power," and bringing a 1GW data center online one month earlier can generate potential benefits of $1 billion to $2 billion. At that scale, paying an extra $37 per MWh is almost negligible.

Fuel cells also have two additional advantages: first, the need for redundant configuration is the lowest, requiring only about 9% excess installed capacity (about 725MW) to serve a 500MW IT load, while CCGT requires as much as 47% redundancy (about 979MW); second, fuel cells directly output DC power, which can seamlessly connect to the 800V high-voltage DC rack architecture promoted by Nvidia, eliminating transformers and inverters and compressing power losses of about 10% to 12% down to about 3%.

Supply Chain Under Severe Strain: From Gas Turbines to Fuel Cells, None Are Spared

Supply tightness in the gas turbine market has reached historic extremes. Goldman Sachs data show that global gas turbine orders reached 100GW in 2025, a sharp jump from 55GW in 2024; momentum was even stronger in 2026, with 67GW signed in the first half alone, including 38GW in the second quarter alone.

The order conditions of major manufacturers are striking: GE Vernova's backlog plus reserved capacity reached 116GW, expected to exceed 125GW before the end of the year, with 2030 capacity fully sold out and more than 50% of 2031 capacity already sold; Siemens Energy's cumulative commitments reached 87GW, of which 24GW is related to data centers, and even including all announced capacity expansions, there will still be a supply-demand gap of about 10% in 2030; Mitsubishi Heavy Industries is already discussing projects for delivery in the 2030s.

Engine manufacturers are similarly overwhelmed: Wärtsilä's 2028 capacity is sold out and it is negotiating slots for 2029 to 2030; INNIO's backlog plus reserved volume exceeds 15GW, more than four times its deliveries over the past 12 months, and it has just signed a 1.1GW baseload power order for a hyperscale data center; Caterpillar's large engine backlog has grown more than 3.5 times; Cummins' orders are booked through 2028; Rolls-Royce said its data center business already accounts for more than 80% of its power generation sales.

The pressure from supply bottlenecks has spread to extreme countermeasures: according to reports, Musk's SpaceX has begun building its own turbine blade factory in Texas to break through the bottleneck; some AI developers have even restarted industrial boilers and steam turbines just to get machines running before 2032.

Fuel cells are also facing capacity constraints. Goldman Sachs estimates that even if Bloom Energy's newly built 2GW production line operates at full capacity, cumulative deliveries by 2030 will be only about 7.7GW, far below its forecast global demand of 18GW. To fill this gap, licensees of Ceres Power—including Doosan, Delta, Weichai and others—would need to expand production on a large scale simultaneously, while Doosan's current annual capacity is only 50MW.

Long-Term Way Out: Small Modular Nuclear Reactors Are the Only Permanent Solution

Goldman Sachs also buried a key judgment in the report: small modular reactors (SMRs) are a viable BTM option, "reliable and relatively economical over the long term," but "due to long investment cycles, most data center investments are unlikely to be realized before 2030."

This timeline happens to align with the internal logic of the entire BTM narrative: gas turbines, reciprocating engines and fuel cells are essentially transitional solutions, built on assumptions of 5- to 7-year turbine queues, capacity-constrained fuel cell production lines, and natural gas prices remaining low over the long term. The only solution capable of providing permanent, scalable, fuel-price-volatility-resistant, zero-emission and truly on-site power is modular nuclear reactors deployed within campuses.

Goldman Sachs' timeline almost naturally connects with SMR developers' plans: gas BTM and fuel cells bear the transitional load before 2030, modular nuclear power gradually scales in the 2030s, precisely catching the window when the first batch of on-site gas assets expires and is up for renewal and the trajectory of natural gas prices is uncertain.