A team at Telehouse was once given until the following Friday to place an order for generators that would reach site two years later. Ozgur Duzgunoglu, the company's head of engineering and design, told the story on a panel at Capacity Europe 2026 in London on 14 October, and it set the tone of the session. Builders now order generators about two years ahead and commit to suppliers before the design is finished, Capacity reported the next day; getting a manufacturing slot has moved ahead of getting the best price.
Slots Before Specifications
Miriam van Kooperen, senior director of supply chain at Greenscale, said capacity now worries her more than cost or lead time. She named high-voltage transformers, gas-insulated switchgear and generator engines as the items in short supply, and said some suppliers already describe themselves as full. Her answer is a preferred vendor list and a fixed cut-off date by which the design is broadly frozen. She has also changed her view on standard designs: she now prefers specifications written around performance rather than a named product, so that an operator can move to another supplier when the first choice cannot deliver.
Duzgunoglu put generator deliveries at up to two years and high-voltage transformers at 12 to 18 months, depending on capacity, and warned that a hyperscaler ordering hundreds of units can take the slots of anyone who stops chasing the factory. Telehouse works from one standard design with flexible sub-packages that it can deploy anywhere from Frankfurt or Paris to Thailand, and names four or five suppliers for each main package instead of one. Getting power in Frankfurt, he added, can take five to seven years.
Nelson Pinto de Magalhaes, director of PMO and procurement at Start Campus, which is building a hyperscale campus in Portugal, said early commitments take care of long-lead equipment; the harder job is getting the equipment, the installers and the commissioning teams into the same window. AI deployments change roughly every six months while even a fast-tracked building takes about 18 months, so at Start Campus procurement leads and design follows. Performance specifications widen the supplier pool, he said, but move the risk onto interfaces and commissioning. Justin Ryburn of Kentik described a U.S. customer that runs generators continuously to bring capacity online before its grid connection is ready, and switches them to backup duty once the grid arrives.
Costs That Outrun the Lease
On the same day Data Center Knowledge reported on JLL's Global Construction Costs Index 2026. It treats annual construction cost escalation of 3 to 6 % as the new baseline, against 2 to 3 % before 2020, and forecasts global tender prices rising by another 3.1 to 4.2 % in 2027. Matthew Giddings of JLL's data center team advises reserving production slots for transformers, switchgear, generators and cooling equipment while negotiating the right to change specifications before fabrication starts. He counts electricians, high-voltage specialists and commissioning teams among the scarcest people on a project.
The Engine Is Booked, the Set Is Not
A slot reserved two years out fixes the engine and its rating long before anyone has drawn the yard. Everything that turns that engine into a working set on a particular site - the container, the exhaust and SCR, the attenuators, the fuel system - gets specified later, after the design freeze, and often against a much shorter programme. A performance specification helps there only if the performance is actually written down: ambient conditions and altitude, the noise limit at the boundary, the emission limits the permit will set, the fuel. On most projects those are known early and can go into the enquiry together with the engine order.
Ryburn's example is worth a second look too. A set bought for standby that runs every day until the grid connection arrives is doing prime duty for that period, with all that means for its rating, its service intervals and its air permit. That is better settled when the slot is booked than when the grid date slips.
Sources: Capacity, Data Center Knowledge