Compute squeeze deepens
OpenAI paused its planned U.K. “Stargate” data‑centre project and Microsoft has taken over the Norway site and is renting capacity there, including tens of thousands of Nvidia chips — moves that reflect shifting datacentre allocations. GPU rental prices have surged recently, with one analysis saying rental rates rose 48% in 60 days, signalling tighter short‑term supply for model training and deployment. (qz.com, finance.yahoo.com, tomtunguz.com)
OpenAI has paused its planned United Kingdom Stargate data-center project, while Microsoft has stepped into a Norway site that had been lined up for OpenAI. (qz.com, finance.yahoo.com) Quartz reported on April 13 that OpenAI put the United Kingdom project on hold because of high industrial power costs and the country’s regulatory environment, even as the company signed a new London office lease for 500 workers. Bloomberg, via Yahoo Finance, reported on April 14 that Microsoft agreed to rent capacity at the Norway campus instead. (qz.com, finance.yahoo.com) The Norway site is in Narvik, above the Arctic Circle, and was announced in July 2025 as OpenAI’s first European Stargate project with Nscale and Aker. That plan called for an initial $1 billion facility with 100,000 Nvidia processors by the end of 2026 and room to expand further. (finance.yahoo.com, finance.yahoo.com) Microsoft’s new Norway deal covers 30,000 additional Nvidia Vera Rubin chips from Nscale, according to Bloomberg’s April 14 report. That is a shift from OpenAI using the site directly to Microsoft renting the capacity itself. (finance.yahoo.com) A data center is the warehouse where artificial-intelligence models are trained and run, and the key machines inside are graphics processing units, or graphics chips, that handle many calculations at once. Stargate is OpenAI’s name for a long build-out of those facilities, aimed at securing power, land, cooling, and chips in one package. (qz.com, finance.yahoo.com) The pressure point now is not only chip manufacturing but also getting live capacity fast enough to serve users and train newer models. Tomasz Tunguz wrote on April 13 that rental prices for Nvidia Blackwell graphics processors rose 48% in 60 days, citing Ornn data, from $2.75 to $4.08 an hour. (tomtunguz.com, edgen.tech) That squeeze has already shown up in service reliability and pricing. Tunguz wrote that Anthropic’s Claude application-programming interface uptime fell below enterprise expectations, while Edgen reported CoreWeave raised prices by more than 20% and lengthened contract terms as spot-market costs climbed. (tomtunguz.com, edgen.tech) OpenAI’s United Kingdom pause also shows that electricity and permits can block projects even when chip partners are in place. Quartz said the company still plans to expand in London, but the data-center build itself is on hold while costs and rules remain unresolved. (qz.com) Microsoft and OpenAI are still closely tied, but this week’s moves show capacity being reassigned site by site instead of arriving on a smooth global schedule. In the short term, the companies with signed power, built halls, and reserved Nvidia inventory are the ones with room to move. (finance.yahoo.com, tomtunguz.com)