Data Center InfrastructureIn AI power, capacity is booked before it is bought.
Data Center Infrastructure/NRG Insights
In AI power, capacity is booked before it is bought.
Data centers are on course to more than double their electricity use by 2030, and AI is the main reason. The gear that turns grid power into cooled, usable compute, from substation transformers to the coolant loops at the rack, now sits in factory queues measured in months or years, and the largest buyers reserve output by contract before they place orders. A US specialist in data-center power and cooling is the case we use to read the queue: how long it is, how output is reserved, and where it should ease over the next five years.
High-voltage equipment of the kind that ties a campus to the grid is now booked in factory queues long before the AI hall it will serve is finished.
本研究简报以英文发布。
Utility-class transformers now take years to arrive. Each new generation of accelerators comes with reference designs that show which power and cooling equipment has been engineered around it. And the largest buyers have begun reserving factory output by contract before they place a single purchase order. In the AI build-out, much of the power and cooling gear behind a data hall now has to be booked before it can be bought.
The public accounts show how long the factory queue has become. One US specialist in data-center power and cooling ended 2025 with a backlog more than twice the size of a year earlier, after a fourth quarter in which it booked nearly three times what it shipped. What this market increasingly sells is a place in line. This article looks at why the slowest power and cooling gear is now reserved before it is bought, how that reservation works, and what changes over the next five years.
Why an AI hall is a different build
Demand is the easy part of the story. The IEA's Energy and AI report puts global data-center electricity use at about 415 TWh in 2024, around 1.5% of the world's power, and projects roughly 945 TWh by 2030 in its base case — growth of about 15% a year, with consumption by accelerated, largely AI, servers rising about 30% a year. In the United States, data centers account for nearly half of all electricity-demand growth to the end of the decade. For equipment suppliers, though, the total matters less than the shape of each new hall.
That shape has changed. Uptime Institute's 2025 survey put the average rack, excluding outliers, at 7.5 kW. A reference design drawn up by a large European electrical group and a leading AI-chip maker, for one of the chip maker's rack-scale platforms, supports well over 100 kW per rack, cooled by liquid through coolant distribution units (CDUs) and direct-to-chip loops. A hall built at that density is a different machine: power per rack rises by an order of magnitude, and heat leaves in liquid rather than air.
Well over 100 kW per rack in a rack-scale accelerator design
Cooling
Mostly air, backed by a chilled-water plant
Liquid to the chip through CDUs, plus heat rejection
What the design is built around
A general-purpose server fleet
One accelerator platform and its published reference design
Conventional versus AI data halls: what is being built. Sources: Uptime Institute 2025 survey (via The Register); a published reference design for a rack-scale accelerator platform (via trade-press coverage).
The queue, item by item
The equipment list for a hall is long, but the queue is set by a handful of items whose factories cannot be expanded quickly, and public data put the electrical ones at the front. Wood Mackenzie's lead-time survey shows where the pressure sits in the US market. In the second quarter of 2025, power transformers took an average of 128 weeks from order, generator step-up transformers about 143 weeks, switchgear about 44 weeks and three-phase pad-mount transformers about 46 weeks. US demand for power transformers stood 116% above its 2019 level, and the firm estimated that demand for the year ran 39% above available supply. The IEA adds that wait times for transformers and cables have doubled in three years, that transformer order backlogs grew more than 30% in 2024, and that the power-transformer price index has risen 1.5 times since 2020.
Item
Lead time
Basis
Generator step-up transformer
About 143 weeks
Wood Mackenzie
Power transformer
About 128 weeks
Wood Mackenzie
Large backup generator (industry norm)
About 70–90 weeks
Data Center Frontier
Three-phase pad-mount transformer
About 46 weeks
Wood Mackenzie
Switchgear
About 44 weeks
Wood Mackenzie
Building-level electrical and cooling gear for a data hall
30–60 weeks
NRG estimate
Order-to-delivery lead times for long-lead power equipment, alongside NRG's estimate for building-level data-hall gear. Transformer and switchgear rows: Wood Mackenzie lead-time benchmarking, US market, second quarter of 2025 (September 2025 executive summary). Generator row: industry norms reported by Data Center Frontier, April 2025. The final row is an NRG estimate, deliberately wide.
NRG's research puts order-to-delivery for the building-level electrical and cooling gear a data hall buys at roughly 30–60 weeks (NRG estimate). That is shorter than the headline transformer numbers, and the gap is not a contradiction. The two-to-three-year figures in public reports describe high-voltage power and generator step-up transformers, the utility-class units that sit at the substation. Building-level gear sits closer to the 44- to 46-week public figures for switchgear and pad-mount units, and JLL's market data, as reported by The Register, show chillers, transformers, switchgear and generators all taking considerably longer than the 28-week average for data-center equipment. Large backup generators sit further out still: Data Center Frontier reported in April 2025 that industry norms had stretched to 70 to 90 weeks. The clocks run at once: a campus that needs its own substation inherits the longest one, and an owner who books nothing early inherits all of them.
Building-level electrical and cooling gear for an AI hall
30–60 weeks
NRG estimate of order-to-delivery for the building-level electrical and cooling gear bought for a data hall. Public benchmarks for comparison: US switchgear averaged about 44 weeks in the second quarter of 2025, power transformers about 128 weeks. Source for public figures: Wood Mackenzie.
Booking before buying
Wood Mackenzie describes developers and utilities forced to secure production slots further in advance, and the largest buyers have turned that pressure into contracts. In late 2025, the same European electrical group signed multi-year supply capacity agreements with two large data-center operators, together worth more than $2bn, which Data Center Frontier reported were structured around dedicated manufacturing lines and inventory rather than one-off purchase orders. For a broad electrical group, reserved output has become a product in its own right, sold ahead of the equipment itself.
Reservation also starts on paper. The reference designs published for each accelerator generation, such as the rack-scale design described above, set out the power and cooling equipment a hall is built around. The US specialist said in late 2025 that its 800 VDC reference architecture was already the basis of design, the document a project's engineers build to, on several large AI-factory projects, ahead of a planned product release.
Liquid cooling now enters a hall as part of an accelerator platform's reference design. Once a project's engineers build to that document, the equipment it names is spoken for long before a purchase order is placed.
What a specialist's order book shows
The specialist's reported results show the same queue. Unlike the European group, whose business reaches well beyond data centers, it has data-center power and cooling as its core business. Net sales were about $10bn in 2025; in the third quarter of 2026 they rose about 60% year on year, with organic orders more than doubling. In October 2026, management described typical order timing as about a year or more ahead and said the company was expanding chiller production capacity sharply.
Metric
Reported
Source
Net sales, full year 2025
About $10bn
Results release, February 2026
Net sales, third quarter of 2026
Up about 60% year on year
Results release, October 2026
Backlog, end of 2025
More than doubled in a year
Results release, February 2026
Book-to-bill, fourth quarter of 2025
Nearly 3x
Results release, February 2026
Typical order timing
About a year or more ahead
Earnings call, October 2026
Chiller production capacity
Being expanded sharply
Earnings call, October 2026
A US data-center power and cooling specialist: public metrics from its own reporting, rounded or summarized. Sources: company results releases (February and October 2026); third-quarter 2026 earnings call. No valuation view is implied.
“When factories are full, the scarce product is a place in the production schedule. The capacity agreements signed in 2025 show buyers paying to hold one.”
— Nathan Research Group, NRG Insights
The five-year outlook: when does the queue shorten?
Demand offers no early relief. The IEA's base case takes data-center electricity to about 945 TWh by 2030 and around 1,200 TWh by 2035, and it estimates that grid constraints could delay about 20% of the data-center capacity planned for construction by 2030. Korea shows the grid side of that constraint at home. Government figures published in 2023 counted 732 data centers planned by 2029, carrying 49,397 MW of power demand, 80.6% of it in the Seoul capital region, and the policy response was to allow supply to be refused to new centers that would overload the local grid and to require a grid-impact review before power-hungry projects proceed. The planning numbers are still moving up. In the provisional demand forecast for Korea's twelfth Basic Plan for electricity, data centers consume 26.5 TWh in 2040, 1.7 times the 15.5 TWh the eleventh plan assumed for 2038, and in June 2026 the climate and energy ministry was reported to be weighing a higher overall forecast still. The IEA puts Japan and Korea together at about 5% of global data-center electricity use, a share it expects them to hold through 2030.
New factory lines are how the queue shortens, and they do not arrive at the same pace across the chain. Lines for building-level electrical and cooling gear come first; high-voltage transformer capacity is slower to add.
Supply is responding. Wood Mackenzie counts nearly $1.8bn of transformer-manufacturing investment announced for North America since 2023 and expects the deficit to ease over the coming years, and the US specialist is expanding its chiller output sharply. Korean makers are part of that response. Exports of large transformers, those above 10,000 kVA, reached $1.04bn in the first ten months of 2025, already above the full-year 2024 total of $946m, and the United States took more than 60% of the value, up from 28% in 2022. Korean makers are now adding transformer lines in the United States itself. The queue will not vanish, but its parts will clear at different speeds.
NRG's view of the next five years follows from that. Building-level electrical and cooling gear is where new factory lines arrive fastest, so it is where lead times should shorten first and early reservation should matter least. High-voltage transformers and the grid connections behind them are slower to add and will stay the binding constraint for longer, which keeps pushing large campuses toward on-site generation, the subject of our companion piece on fuel cells for AI data centers. Power architecture is the wild card: 800 VDC power systems are expected to reach customer validation and then deployment later this decade. A change of architecture means new reference designs, and with them a new round of reservations for the equipment those designs specify.
Path
What would have to be true
Where the queue binds
The queue persists
Grid delays and transformer deficits outlast new factory capacity, and AI build plans hold.
Factory output across the chain; reserving it early stays the norm.
The queue clears
New lines for building-level electrical and cooling gear catch up while build plans plateau.
Mainly high-voltage transformers and grid connections; building-level gear returns to ordinary lead times.
The architecture resets
800 VDC power and megawatt-class racks move from validation to volume.
Equipment for the new architecture, as reference designs and basis-of-design documents are rewritten.
Illustrative paths for 2026–2030 — analytical constructs, not forecasts. Each path changes where the queue binds. No valuation or price view is implied.
The lesson travels beyond data centers. Whenever demand runs ahead of factory capacity, in grid equipment, in chipmaking tools, in ships, buyers start paying for certainty of date, and a supplier's real asset becomes its place in the customer's plan. In AI data centers the public record shows that place being secured earlier each year: in reference designs drawn up before the chips ship, and in capacity agreements signed before the purchase orders.
How this article was built. Figures with a cited source are public: company results and earnings calls, the IEA's Energy and AI report, Wood Mackenzie's supply-chain survey, government statistics and press reports. Companies are described rather than named, and figures from their own reporting are rounded or summarized. Figures marked NRG estimate come from NRG's own research; they are rounded into deliberately wide ranges and published only where public data point the same way. Where the text describes industry practice, it is NRG's own analysis of its research and the public evidence cited here, written as general practice; it describes no specific company, project or transaction. 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 large the order books have grown and the IEA shows how large the demand is, but neither shows the queue from the inside: how long-lead equipment is planned, reserved and delivered. That knowledge sits with the people who design, make, buy and connect AI halls. 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 operators and developers
Site selection, power procurement and phasing of AI campuses.
Electrical and cooling equipment makers
How electrical and cooling equipment is specified, built and delivered across the industry.
Grid and utility planners
Interconnection queues, substation build times and Korea's review of large new loads.
Power-distribution engineers
Electrical design for high-density data halls.
Supply-chain analysts
Bottlenecks in electrical steel, copper and power semiconductors.
Data-center load forecasts and the demand assumptions behind national power plans.
项目怎么推进
1范围界定
把您要研究的行业问题，转成具体的专家画像和问题清单，对准您需要作出的决定。
2寻访与核实
每一位专家在访谈前都要过一遍：课题相关性、经验的时效性，以及有没有利益冲突。
3组织与归纳
按您的时间安排访谈；需要的话，另附一份按问题逐条对应的书面归纳。
如果您的团队正在评估 long-lead power and cooling equipment、the AI data-center power and cooling chain，乃至范围更大的 data-center infrastructure market，请告诉我们您眼下要做的是什么决定。
