Inference Moves Towards Cities

AI data-center construction is moving closer to cities as companies spend for faster AI responses, not just giant remote training clusters. (businessinsider.com) DataBank said on April 21 it secured a $2.0 billion construction loan to build the first three data centers on a new campus in Red Oak, Texas, about 20 miles south of Dallas. The company said the first phase will deliver 180 megawatts of capacity, with Mitsubishi UFJ Financial Group leading the bank group. (databank.com) Business Insider reported MUFG is also leading a separate effort to raise about another $600 million for a fourth building at the same Red Oak campus. The site sits outside Dallas rather than in a remote power market, reflecting demand for AI facilities nearer large user bases. (businessinsider.com) Inference is the part of artificial intelligence that answers a prompt after the model has already been trained. Google Cloud calls it the moment a model “stops learning and starts working,” and NVIDIA says strong inference performance depends on low latency, or very short delays. (cloud.google.com, nvidia.com) That changes where capacity gets built. Google Cloud says online model deployment attaches physical resources so a model can serve low-latency requests, and Amazon Web Services says edge AI pushes inference closer to users in bandwidth-sensitive and globally distributed applications. (cloud.google.com, docs.aws.amazon.com) DataBank has spent years positioning itself for that pattern. On its edge-strategy page, the company says it targets dozens of large metros and can place workloads within 50 miles of half the U.S. population. (databank.com) The Dallas loan follows a separate capital raise from October 2024, when DataBank announced roughly $2.0 billion in new equity led by a $1.5 billion investment from AustralianSuper. The company said that money would fund more than 850 megawatts of additional capacity across its portfolio. (databank.com) DataBank has kept adding financing on top of that. In September 2025 it announced a $1.1 billion hyperscale asset securitization, and in October 2025 it said it had expanded another financing vehicle to $1.6 billion for ongoing and future construction. (databank.com, databank.com) Other operators are building around the same idea. Akamai said last month it had expanded its cloud footprint to 41 data centers as it pushed further into AI inference, tying that growth to distributed infrastructure for end-user performance. (akamai.com) DataBank said two weeks ago that its platform now includes more than 70 high-performance-computing-ready data centers in more than 25 U.S. markets. The Red Oak project adds another large block of capacity near a major metro, where the AI buildout is increasingly aimed at serving live prompts instead of only training the next model. (databank.com)