Fuel Cells & On-site PowerFuel cells sell AI data centers time, and US buyers are signing up.
Fuel Cells & On-site Power/NRG Insights
Fuel cells sell AI data centers time, and US buyers are signing up.
AI data centers have run out of patience with the grid. Large hookups can take up to seven years, and a leading gas-turbine maker has several years of its current output already contracted or reserved. Solid oxide fuel cells, which can be on site in months, have become one of the fastest routes to power, and in the United States, two buyers alone, a utility and a cloud provider, have frameworks for close to 4 GW of them. Korea has the same technology and fuel-cell factories that can make far more than its national auction buys, but that auction was built for small distributed plants, not data centers. A US fuel-cell maker that sells in both markets is the case we use to read them.
For a new AI data center, how soon power can be on site has become a deciding factor, and on-site generation is one way to shorten the wait.
本研究简报以英文发布。
In the United States, the question of whether fuel cells can power AI data centers is being settled in contracts. One US maker of solid oxide fuel cells plans to double its factory capacity by the end of 2026, and a single framework with one cloud provider allows for close to 3 GW.
One variable runs through the filings and the buyers' announcements: how soon power can be on site. In the United States, that variable has moved a specialist product into framework agreements sized in gigawatts. Korea, that maker's second-largest market, shows the same technology from the other side: plants that can build far more than the national auction buys, and no buying route shaped for data centers.
The orders came first — in America
The order book was built quickly and in public. In late 2024, a large US utility agreed to secure up to 1 GW of the maker's solid oxide fuel cells for data centers and other large users; the IEA, which lists the deal in its Energy and AI report, notes that these fuel cells were bought to run as base load. In the spring of 2026, a major cloud provider expanded its partnership with the maker to close to 3 GW, of which more than 1 GW was already contracted at announcement. Between them, the two frameworks allow for close to 4 GW; both figures are ceilings.
The results followed. The maker's revenue rose by more than a third in 2025, and its product backlog more than doubled from a year earlier; the company credited significant growth from the AI data-center industry. By the third quarter of 2026, quarterly revenue was more than twice its level a year earlier. The company's own datasheet, dated early 2026, put cumulative deployments at about 1.5 GW across more than a thousand sites — so the cloud provider's framework alone is larger than everything the maker had installed up to that point.
One buyer's framework ceiling against the maker's installed base
≈3 GW vs ≈1.5 GW
One cloud provider's framework allows for close to 3 GW of the maker's fuel-cell systems, more than 1 GW of it contracted at announcement in 2026. The maker's early-2026 datasheet put cumulative deployments at about 1.5 GW. The first figure is a ceiling on future purchases, the second everything installed to date. Sources: company release and datasheet.
Behind the gigawatts are thousands of units. The maker's annual report describes its platform as made of repeatable building blocks, from printed cells to stack assemblies and columns. At roughly 300 kW a unit, the cloud provider's initial contracted tranche of more than 1 GW alone works out to well over 3,000 units, each of which has to be built, shipped and installed.
The queues behind the orders
The reason for the orders sits in the queues. The IEA expects global data-center electricity use to more than double, from about 415 TWh in 2024 to around 945 TWh by 2030, and estimates that grid constraints could delay around 20% of the data-center capacity planned for construction by then. In the United States, the Department of Energy put data centers at about 4.4% of national electricity use in 2023 and projects 6.7–12% by 2028. Where the load clusters, the queue is longest: in Virginia, a major utility warned in 2024 that connecting large data centers could take up to seven years.
Building generation on site is the obvious workaround, but the conventional route has its own queue. A leading gas-turbine maker reported gas-power equipment backlog and slot reservations of more than 100 GW in the third quarter of 2026, several times what its factories turn out in a year, with plans to raise that output by about half by 2030. At the current rate, that is several years of production spoken for. Fuel cells take a shorter path: the fuel-cell maker's annual report says on-site generation can, in some project scopes, deliver power within about three months, and the company said it delivered a fully operating system to the cloud provider in under two months.
Route
What sets the wait
Public marker
Utility grid connection
Interconnection studies, substations and new transmission
Up to seven years for large hookups in Virginia; the IEA sees around 20% of planned capacity at risk of delay
On-site gas turbines
Turbine factory slots
A leading maker's backlog and reservations equal several years of its annual output
On-site solid oxide fuel cells
Project scope
Power within about three months in some scopes; one system delivered fully operating in under two months
Three routes to power for a new AI data center, and what sets the wait on each. Sources: Data Center Knowledge; IEA; company filings and releases.
Solid oxide cells run hot, which, the US Department of Energy notes, results in a slow start-up; they suit steady running — base load, as with the units in the utility framework, or a bridge until a grid connection arrives — more than standing idle as emergency backup.
The conventional fix for a slow grid is a gas plant on site, but its turbines come off factory lines with years of orders already ahead of them. A fuel-cell system that can be running within months sells a data center time as much as power.
Korea: the same technology, the opposite constraint
Korea has fuel-cell factories and a long operating record. It is the US maker's second-largest market, where it began commercial operations in the late 2010s through distribution partnerships with affiliates of a Korean conglomerate; a joint venture between the two makes SOFCs in Korea. A Korean fuel-cell maker that built its business on a different cell chemistry completed its own SOFC plant in mid-2025; early cell yields were low, and the company has since reported steady improvement toward its targets on core processes. Its first commercial SOFC fleet, a project of under 10 MW, was scheduled to begin commercial operation in the second half of 2026. Across the main makers, annual fuel-cell production capacity is about 675 MW.
Demand, however, is set by the state. New fuel-cell power plants are bought through auctions under the Clean Hydrogen Portfolio Standard (CHPS), whose general-hydrogen market opened in 2023 mainly for fuel cells and distributed generation. In the 2025 general-hydrogen auction, 77 plants bid for 3,137 GWh and 52 won 1,355 GWh; every winner was 20 MW or smaller, 88% of the awarded volume sat in the Seoul metropolitan area and other large cities, and commercial operation starts around 2027. The market is shrinking: one Korean fuel-cell maker's investor presentation puts it at 175 MW, 175 MW and 179 MW for 2023–2025 and expects about 125 MW in 2026, when the auction offers 930 GWh of general hydrogen power alongside 500 GWh of clean hydrogen power.
Korea's fuel-cell factories against its auction
≈675 MW vs ≈200 MW a year
Annual production capacity of Korea's main fuel-cell makers against the roughly 1,300 GWh (about 200 MW) of general-hydrogen power auctioned each year in 2023–2025. Even at that pace, Asia Economy notes, utilization would stay under 30%. Source: Asia Economy, May 2026.
The pressure that created the American market is building in Korea too. Data-center applications for grid-impact assessments total about 78.7 GW, including duplicate and speculative requests, about 44 GW of it in the Seoul metropolitan area; AI data centers have requested about 8.4 GW of specific supply. A new transmission line typically takes around nine years. The Distributed Energy Act, in force since June 2024, lets producers and consumers in designated zones trade power directly. What Korea lacks is a buying route shaped for data centers: the auction is aimed at distributed generation and in 2025 went to plants of 20 MW or less spread across many sites.
Transmission is the slowest part of the grid to build: in Korea, a new line typically takes around nine years, while data centers have already asked for gigawatts of supply. The fuel-cell factories exist; what is missing is a buying route shaped for data centers.
NRG's reading of these figures is that Korea's constraint sits on the buyer's side of the factory gate: the plants exist, and the auction buys well under what they can make, while the buying route and grid rules set the pace.
Question
United States
Korea
Who sets demand
Individual buyers under framework agreements: a utility, up to 1 GW; a cloud provider, close to 3 GW
Annual government auctions for hydrogen power
Typical commitment
Close to 3 GW from one buyer, more than 1 GW of it contracted at announcement
Plants of 20 MW or less; 1,355 GWh awarded in 2025
Factory position
The US maker plans to double its factory capacity by end-2026
About 675 MW a year of capacity against about 200 MW auctioned
One technology in the US maker's two largest markets. Sources: company filings and releases; press coverage; Korea's Ministry of Trade, Industry and Energy; Asia Economy.
The five-year outlook, 2026–2031
Published forecasts for the solid oxide market point up, though their scopes differ. MarketsandMarkets and BCC Research both put the entire market near $3bn in 2025, and both project strong growth to 2030.
Source
2025 base
Forward figure
Annual growth
MarketsandMarkets (market)
$2.98bn
$11.61bn by 2030
31.2%
BCC Research (market)
$3.2bn
$9.6bn by 2030
24.4%
Solid oxide fuel-cell market forecasts. Forecast scopes differ and BCC's release gives inconsistent base-year figures; treat both as approximate. Sources: MarketsandMarkets; BCC Research.
The demand signals point the same way. A survey by the US fuel-cell maker found nearly two in five data-center facilities expected to use primary on-site generation by 2030, up from about one in eight in the previous survey, and more than a quarter expecting to run fully on site, up from almost none. A vendor's survey deserves caution, but its direction matches the IEA and US Department of Energy projections above.
Year
Scheduled or projected
Source
2026
The US fuel-cell maker's planned doubling of factory capacity is due by end-2026; Korea's auction offers 930 GWh of general hydrogen power
Company annual report; Asia Economy
2028
A leading gas-turbine maker expects higher annual turbine output; the DOE projects data centers at 6.7–12% of US electricity use
Company results; US Department of Energy
2030
The same turbine maker aims to lift annual output by about half from today's level; the IEA projects data-center electricity use of around 945 TWh, with around 20% of planned capacity at risk of grid delay
Company results; IEA
What the public record already schedules or projects through 2030. Every entry comes from a company or government source, not an NRG estimate. Sources: company filings and results; Asia Economy; US Department of Energy; IEA.
Path
What would have to be true
What it means for fuel cells
Fuel cells scale with AI data centers
Framework ceilings turn into installed capacity, and fleets perform as specified in continuous duty
Fuel cells become a standard layer of AI data-center power
The grid catches up first
Transmission and turbine capacity arrive faster than expected
Fuel cells settle into a bridge role, and growth slows after the first wave
Field performance falls short
Fleets in continuous duty perform below their specifications
Buyers lean back toward turbines and the grid, and service costs rise
Three paths to 2031 — analytical constructs to frame the question, not forecasts. No view on any company's results or securities is implied.
Three signals will show which path is unfolding: whether framework ceilings turn into contracted, installed capacity; whether fleets in continuous data-center duty perform as specified, as far as public disclosures show; and, in Korea, whether auction volumes and distributed-energy rules open a buying route sized for data centers rather than for small plants.
“In America, two buyers alone have frameworks for close to 4 GW; in Korea, the national auction buys less than a third of what the main fuel-cell factories can make. It is the same technology in two very different markets.”
— Nathan Research Group, NRG Insights
The lesson travels. When a power source sells time, its market forms where buyers are short of time and the rules let them act on it. American data-center buyers have both today. Korea has the factories; what it lacks, for now, is a way for a data-center buyer to use them.
How this article was built. Every figure here comes from a public source — company filings and releases, the IEA, the US Department of Energy, Korean government releases, a Korean fuel-cell maker's investor presentation and press reports. Companies are described rather than named, and figures specific to a single company are rounded or generalized. This article publishes no NRG estimate; NRG Insights uses those rounded, deliberately wide ranges from NRG's own research only where public data point the same way. NRG's own research shaped the framing and how the public figures are read. This is business and market research, not a view on any security.
与 NRG 合作的方式
与 Nathan Research Group 合作。
Filings show how fast order books have grown and auction results show how much Korea buys, but neither shows how a data center actually chooses, builds, permits and runs its on-site power. That knowledge sits with the people who design, buy, permit and operate these systems. Reaching them, compliantly, is what we do.
韩国首家专家网络 — 首尔，2013年至今
我们会请谁到场
Data-center power and energy leads
Owner and developer teams who choose between grid, turbines, fuel cells and batteries for new AI data centers.
Utility and interconnection specialists
Large-load connection queues in the US and Korea's grid-impact review for new data-center demand.
Fuel-cell engineers
Stack and system design and reliability.
Manufacturing scale-up specialists
Ramping high-volume production lines.
Gas supply and permitting specialists
Pipeline capacity, air permits and local siting rules for on-site generation.
