Inference moves to cities
- Latency‑sensitive AI inference is driving data centres closer to urban users rather than remote mega‑campuses. - DataBank raised $2 billion to expand urban and near‑urban facilities aimed at inference workloads. - Urban facilities shift the competitive edge to operators who can secure city power, permits and advanced cooling. (businessinsider.com)
DataBank has lined up $2 billion to build three new data centers south of Dallas, betting AI work is moving closer to cities. (prnewswire.com) The loan, announced April 21, will fund the first three of eight planned buildings on DataBank’s Red Oak, Texas, campus. The first phase is already leased and totals 600,000 square feet and 180 megawatts of power. (prnewswire.com) DataBank said the financing should pull construction and delivery forward by about 18 months. MUFG Bank led the deal, which the company called its largest construction financing so far. (prnewswire.com) The shift starts with the difference between training and inference. Training builds an AI model in giant clusters, while inference is the live step that answers a user’s prompt, labels a camera feed, or powers a software agent in real time. (mckinsey.com) Inference rewards shorter network trips. McKinsey said in December 2025 that inference workloads were accelerating data-center buildouts in metro and near-metro areas optimized for low round-trip time and dense network links. (mckinsey.com) That is colliding with a power market that already favors operators with grid access in big metros. Jones Lang LaSalle said in its 2026 outlook that power, not land price or location, had become the primary site-selection criterion because utilities were imposing multiyear waits for new connections. (jll.com) The result is a different map from the first AI building wave. Instead of only remote mega-campuses for model training, companies are adding smaller regional sites for fast-response computing closer to users and enterprise customers. (mckinsey.com; edgeir.com) DataBank has been positioning for that market for more than a year. In October 2024, it said it had raised nearly $2 billion in equity, including $1.5 billion from AustralianSuper, to expand AI-focused capacity and a 480-megawatt Red Oak campus while tripling power across major U.S. metro markets. (databank.com) Other infrastructure players are making the same urban and distributed bet. NVIDIA said in March that AT&T, T-Mobile, Comcast and Charter’s Spectrum were building “AI grids” on telecom facilities to run inference closer to users, devices and data. (nvidia.com) A separate effort from Electric Power Research Institute, NVIDIA, Prologis and InfraPartners is planning 5- to 20-megawatt inference sites near utility substations, with at least five U.S. pilots targeted by the end of 2026. (edgeir.com) The race now looks less like a land grab in remote counties and more like a contest for urban power, permits, fiber routes and cooling systems that can handle denser racks. DataBank says its Red Oak build is aimed at getting that capacity into the Dallas market faster. (jll.com; prnewswire.com)
