OpenAI signed a lease for 88,500 square feet at King's Cross in the week to 13 April 2026, framing the site as its biggest research base outside San Francisco and planning for more than 500 UK staff. In the same week the company paused its Stargate UK data-centre buildout. The pause was attributed in OpenAI's public statements to UK industrial electricity prices running at roughly four times US rates and close to twice French levels, with grid-connection delays of three to eight years against 18-24 months in comparable jurisdictions. The company said it would restart Stargate UK when regulation and energy costs permit long-term infrastructure investment.
The geographic split matters because The UK Sovereign AI Unit's policy design assumes research and training compute can grow together on British soil. OpenAI's decision fixes the opposite pattern: research in London, training compute in Texas. Frontier model training runs now consume GPU-hour volumes that make co-location of researchers and training clusters a commercial precondition, not a preference. A King's Cross office serves product and research iteration; it does not serve pre-training at frontier scale. The UK will recruit researchers; the training will happen where the grid can carry it.
A 3-to-8-year grid-connection queue is a binding constraint on UK sovereign compute that neither the Sovereign AI Unit's equity fund nor its separate £250m cloud procurement can compress. National Grid ESO connection timelines are set by network capacity and planning law, not by fund scale, and the Sovereign AI Unit is not a grid-investment instrument. That gap is the gap the fund has not yet addressed in public. DSIT's response has been to fund an AI Research Resource allocation per investee; OpenAI's response has been to lease an office and site the compute elsewhere.
A 500-staff London research office at King's Cross brings senior engineering roles, research output and a local tax base, and OpenAI's expanded UK headcount is not in itself a loss of sovereignty. The difficulty is that the Sovereign AI Unit was designed on the premise that British frontier-model capacity could be built on British compute, and a US lab's decision to pause its British compute buildout this week puts that premise under strain. The unit's infrastructure cohort will operate alongside a US lab that has already chosen the alternative answer to the grid-cost question.
