Compute Buildouts Face Friction

OpenAI has paused its Stargate data center project in the United Kingdom after saying the country’s energy costs and regulatory conditions do not support long-term infrastructure investment. The pause was reported on April 9, 2026, and it cuts against Britain’s push to attract large artificial intelligence campuses. (cnbc.com) That sounds like a company-specific setback, but it points to a basic fact about artificial intelligence: the smartest model in the world still needs a physical building full of chips, cables, cooling gear, and huge power contracts. A modern training cluster is less like a software launch and more like trying to open an aluminum smelter next to a substation. (capacityglobal.com) OpenAI’s original Stargate plan was presented in January 2025 as a project that could spend up to $500 billion on artificial intelligence infrastructure. Once numbers get that large, the question stops being “where do we want to be?” and becomes “where can we actually get power, permits, and land on time?” (telegraph.co.uk) Britain is a hard place to answer that question because industrial electricity is expensive and grid connections can be slow. Recent reporting on the Stargate pause said United Kingdom industrial power prices were about four times higher than in the United States, which is a brutal handicap for a facility that turns electricity directly into model training. (thenextweb.com) Regulation is the second brake. OpenAI said it would move forward only when the “right conditions” exist, and multiple reports tied that to unresolved rules around building, energy, and artificial intelligence policy, including copyright questions that matter to model developers. (bloomberg.com, thenextweb.com) So the map of artificial intelligence buildouts is starting to look less like a map of talent hubs and more like a map of cheap, reliable electricity. That pushes new data halls toward places with spare generation, easier transmission access, and fewer bottlenecks, even if those places are far from London, San Francisco, or other expensive metros. (oilprice.com) At the same time, demand for chips has not cooled at all. On April 6, 2026, Anthropic said it expanded its use of Google Cloud’s Tensor Processing Units, with multiple gigawatts of capacity, while keeping Amazon Web Services as its primary cloud and training partner. (anthropic.com) A gigawatt is the kind of number usually used for power plants, not software companies. CNBC reported Anthropic’s deal gives it access to about 3.5 gigawatts of computing capacity tied to Google’s artificial intelligence processors, with Broadcom producing future chip generations. (cnbc.com) That is why these two stories fit together. One company is discovering that the wrong location can stall a giant project, while another is locking up supply years ahead so it does not get stranded without enough hardware when new models need to be trained. (datacenterknowledge.com) Anthropic is also pushing custom-chip work, which shows how control is shifting from just buying graphics processors to shaping the full stack of silicon, networking, and cloud access. If power determines where the building goes, chip partnerships increasingly determine whether the building is useful when it opens. (en.sedaily.com, cnbc.com) The result is a stranger artificial intelligence race than people expected in 2023. The bottleneck is no longer just better models or more money; it is whether a company can secure transformers, transmission, permits, cooling equipment, and custom chips in the same place at the same time. (capacityglobal.com, anthropic.com)