The pace of AI data center expansion is currently constrained not by capital, but by electricity. Orders for chips and server rooms can be locked in ahead of time, yet the time required to deliver power to campuses keeps lengthening 鈥?in the United States and Europe, queue times for connecting to the public grid will continue to stretch over the next five years. As power availability replaces capital expenditure as the hard constraint on deployment, self-built, self-used "behind-the-meter" solutions that avoid queuing have been pushed to the forefront.
According to information from Zhufeng Trading Desk, the team led by Michele Della Vigna at Goldman Sachs published a report on September 23 titled "Carbonomics: Behind-the-meter power solutions for data centers: gas turbines, fuel cells and reciprocating engines," which provides a framework that scores gas turbines, fuel cells, and reciprocating engines side by side, and raises the 2030 behind-the-meter power installed capacity forecast from 40GW to 67GW. What supports this upward revision is a set of faster electricity consumption readings.
There are two conclusions: all three technology routes will be used; after reweighting by delivery speed and availability, fuel cells rank first in the composite score.
Why the behind-the-meter power pie suddenly became larger
Demand-side readings changed first. Global data center capacity is expected to reach 217GW by 2030, up from 101GW in 2025; U.S. data center demand will reach 108GW over the same period. Based on this, global data center electricity demand in 2030 will grow 170% compared with 2025, with more than 60% of that coming from the United States. This growth rate is higher than previous assumptions and directly collides with the grid's delivery capability.
The grid's delivery capability is deteriorating. The median time from U.S. grid interconnection application to commercial operation has approached five years for projects completed in 2023鈥?024; new high-voltage transmission lines have fallen from an annual average of about 1,700 miles in 2010鈥?014 to an annual average of 350 miles in 2020鈥?023. Transmission and generation facilities usually take five to ten years to build, and planning processes cannot keep up with this wave of incoming demand.
With upward demand revisions and a slower grid combined, the conclusion on the behind-the-meter side is this forecast upgrade: behind-the-meter generation installed capacity rises from 40GW to 67GW, and all three routes are indispensable.
Where each of the three routes is stuck
The bottleneck for conventional gas has shifted from technological maturity to production capacity and lead times. Under normal market conditions, large gas projects take about two to three years from order to commissioning; today, delivery cycles for combined-cycle gas turbines are four to seven years, simple-cycle is 18 to 36 months, and heavy-duty gas turbines are generally extended to about five to seven years. Manufacturers' capacity expansion is not enough to immediately change the tight balance: GE Vernova's contracted volume will exceed 125GW by year-end, its 2030 capacity is already sold out, and more than half of its 2031 capacity slots have also been pre-sold; Siemens Energy's backlog is close to 69GW, with another 27GW reserved, implying a lead time of about three to four years; projects under discussion at Mitsubishi Heavy Industries will be scheduled for delivery in the 2030s.
Fuel cells' strength is speed of deployment and availability, not generation cost. Modular projects have a delivery window of about 12 to 24 months, do not need to be connected to the grid, have the highest full-load operating hours, and require the least additional backup capacity.
This scoring system adds together delivery speed (20%), availability (15%), levelized cost of electricity (15%), and initial capital expenditure (10%), with more than 60% of the weight placed on execution rather than economics. As a result, fuel cells rank first with 76.6 points, aeroderivative gas turbines score 68.2 points as the most balanced conventional option, reciprocating engines score 67.0 points, and heavy-duty gas turbines and combined cycle score 59.6 points. The scoring does not hide shortcomings: fuel cells' levelized cost of electricity is about US$117/MWh, roughly 45% higher than combined cycle and reciprocating engines.
How large the fuel cell opportunity is and who can capture it
According to calculations in the report, fuel cells correspond to about 12GW of installed capacity in the United States and about 18GW globally by 2030, equivalent to a cumulative equipment market of about US$35 billion (United States) and US$55 billion (global). Gas behind-the-meter plus fuel cells combined can cover about 28% of U.S. and about 25% of global data center electricity demand in 2030, while in 2025 that proportion was close to zero.
The supply-side gap is even clearer. Based on full production and 85% utilization, Bloom Energy's 2GW expansion can cumulatively deliver about 7.7GW of solid oxide fuel cells by 2030, far below the global installed capacity forecast of about 18GW; to realize the forecast, multiple manufacturers need to expand production simultaneously, including Ceres Power's licensees Doosan, Delta Electronics, and Weichai Power.
Beneficiaries are divided by route: on the fuel cell side are Ceres Power, Weichai Power, and Delta Electronics; on the gas behind-the-meter side are GE Vernova, Siemens Energy, Mitsubishi Heavy Industries, and INNIO.
What to watch next
For outcome indicators, watch manufacturers' quarterly disclosures: orders, backlogs, and capacity sold-out years for gas turbine and fuel cell manufacturers, involving GE Vernova, Siemens Energy, Mitsubishi Heavy Industries, as well as Ceres Power and its licensees Doosan, Delta Electronics, and Weichai Power.
For mechanism indicators, watch three things: whether procurement and grid-connection announcements for data center captive power plants continue to increase, whether gas turbine delivery cycles continue to lengthen, and whether fuel cell unit costs and delivery cycles decline as forecast.
Disconfirming conditions: if public grid expansion significantly shortens interconnection queues, or if fuel cell cost declines fall short of expectations and delivery cycles instead lengthen, forecasts for behind-the-meter power share and fuel cell share would both need to be revised downward.
Verification window: October 15, 2026 to November 15, 2026, namely the third-quarter earnings window, as well as the disclosure dates of subsequent captive power procurement announcements.
The power bottleneck is reshaping the order of technology selection, and the order of technology selection is reshaping who wins this round of orders. The move from 40GW to 67GW is an upward revision, not already realized capacity; what to watch next is whether lead times and orders continue to tilt toward the execution side, and whether the fuel cell cost curve declines as forecast.
Comments