Hyperscale
Massive data centres engineered to scale efficiently for large computing workloads.
Power access now decides where hyperscale campuses get built. Britain's connection register held 50 GW of data-centre demand in February 2026, above the 45 GW national peak, and OpenAI paused its Cobalt Park site in North Tyneside on 23 April.
Last refreshed: 1 October 2026 · Appears in 1 active topic
Have AI training requirements made the original hyperscale design obsolete?
Timeline for hyperscale
Mentioned in: OpenAI pauses Cobalt Park Stargate site
Data Centres: Boom and BacklashMentioned in: Maine passes first US statewide DC freeze
Data Centres: Boom and BacklashIEA: 17% growth, $700B capex run-rate
Data Centres: Boom and BacklashMentioned in: Blackstone £10B Blyth, Amazon €33.7B EU
Data Centres: Boom and BacklashStargate US: $500B headline, 1.2 GW operational
Data Centres: Boom and BacklashBackground
hyperscale describes data centres that a single cloud or technology company builds and runs for its own workloads, rather than leasing space to many tenants. The facilities use standardised, modular halls and dense server racks, and a campus commonly carries 100 MW or more of IT load. At that size, electricity supply, transmission access and land decide where capacity can go.
The model grew out of public cloud computing and content delivery, where concentrating infrastructure cuts the cost of each unit of compute. AI training has pushed power needs sharply higher. The International Energy Agency reported on 16 April 2026 that global data-centre electricity use rose 17% in 2025, with AI-focused facilities up about 50%.
The term matters because it names both a design and an operating model, and both now collide with public systems built for smaller loads. One campus can ask a grid for as much power as a city, so planning authorities, grid operators and state legislatures increasingly treat hyperscale growth as an energy policy question as much as an industrial one.
Grid queues now gate hyperscale sites
A hyperscale campus needs a grid connection sized like a town's, and in Britain that connection is now the scarce input. By February 2026 The National register held about 140 data-centre applications asking for 50 GW in total, more than the 45 GW peak the whole country drew on 11 February. Many of those offers are speculative, but even the real ones exceed the transmission headroom on offer this decade.
The first casualty came on 23 April, when OpenAI paused its planned Stargate campus at Cobalt Park, blaming high energy costs and an "unfavourable regulatory environment". Blackstone's and Nscale's builds in the same AI Growth Zone carried on, which shows capital choosing between UK sites rather than leaving altogether.
Hyperscale builders turn to their own generation
Operators that cannot wait for transmission are building their own power. xAI won approval in March 2026 for 41 gas turbines, 1.2 GW in all, at its Colossus complex outside Memphis. Meta's El Paso campus is planned to run on a dedicated 366 MW gas plant, which still needs approval from the Public Utility Commission of Texas, and the Wonder Valley campus in Alberta carries a 7.5 GW gas plan.
That route has its own queue. GE Vernova's gas-turbine backlog stood at 80 GW in December 2025, with deliveries booked into 2029, and the IEA expects 15 to 27 GW of onsite gas to power US data centres by 2030. A turbine ordered now also commits its owner to a fossil asset for decades.