AI's infrastructure rethink
Big AI players are recalibrating how they scale compute as the capital, political and logistical costs of new data centres become clearer. OpenAI has paused its 'Stargate UK' data‑centre plan, a sign that building massive new facilities is politically and financially fraught, while Alphabet is leaning into expanded AI‑chip partnerships with Intel instead of—or alongside—big new buildouts. (engadget.com) (ibtimes.com.au)
OpenAI just hit pause on a British data-centre project it announced with Nvidia and Nscale in September 2025, saying the United Kingdom’s industrial power prices and regulatory hurdles make the math too hard right now. (engadget.com) (cnbc.com) At almost the same moment, Google went the other way on the same problem: it expanded a multiyear deal with Intel so Google Cloud can keep using Intel Xeon 6 processors and co-develop custom infrastructure chips instead of betting only on giant new building sprees. (intel.com) (cnbc.com) A data centre is just a warehouse full of computers, but an artificial intelligence warehouse also needs huge amounts of electricity, cooling water, land permits, grid connections and specialized chips. OpenAI’s UK pause shows that the bottleneck is no longer just getting graphics processors from Nvidia. (cnbc.com) (engadget.com) Bloomberg reported that OpenAI pointed to energy costs and policy uncertainty, and one report said British industrial electricity can run about four times higher than in the United States. That turns every training run into a utility bill problem before it becomes a software problem. (bloomberg.com) (thenextweb.com) Google’s Intel move points to a different lesson: if new campuses are slow and expensive, squeeze more work out of the machines you already have. Intel said the partnership is about “balanced systems,” with general-purpose central processors and network offload chips working alongside artificial intelligence accelerators. (intel.com) (reuters.com) That matters because artificial intelligence jobs do not run on graphics chips alone. Reuters said shifting artificial intelligence use is reviving demand for central processing units, and CNBC reported Intel’s Xeon 6 chips will handle both training and inference work inside Google’s infrastructure. (reuters.com) (cnbc.com) The old story was simple: bigger models need bigger clusters, so the winners would just pour more concrete. The new story is messier: power markets, planning rules, copyright fights, packaging technology and chip mix now shape how fast artificial intelligence can grow. (engadget.com) (thenextweb.com) (intel.com) That is why these two announcements fit together. OpenAI’s pause says a promised site can still stall when power and politics do not line up, while Google’s Intel deal says the next race is also about redesigning the plumbing inside the warehouse you already have. (cnbc.com) (techcrunch.com) For governments, this is a warning that “build here” is not enough without cheap power and faster approvals. For chipmakers, it is a sign that selling the picks, shovels and pipe fittings of artificial intelligence may be easier than waiting for every customer to break ground on a new megaproject. (datacentrereview.com) (intel.com)
