AI infrastructure moves upstream
Recent coverage highlights a shift: investors and corporates are funneling capital into the physical stack—data centers, power, land and deployment certainty—rather than only models or apps, with large capex rounds and project consolidations reported. (techstartups.com)
Artificial intelligence money is moving out of chatbots and into the physical bottlenecks: power contracts, data centers, grid access and land. (bloomberg.com) Oracle said on April 13 it plans to buy up to 2.8 gigawatts of Bloom Energy fuel-cell capacity for artificial intelligence and cloud sites, with an initial 1.2 gigawatts slated for 2026 and 2027 projects in the United States. (bloomenergy.com) In Ohio, SB Energy, a SoftBank Group company, said in March it would build a 10-gigawatt artificial intelligence data center at the Department of Energy’s Portsmouth site and pair it with 10 gigawatts of new generation, including 9.2 gigawatts of natural gas. (energy.gov, datacenterdynamics.com) The constraint is no longer just graphics chips. A data center is a warehouse of servers, and it cannot open without transformers, switchgear, permits and enough electricity to run around the clock. (bloomberg.com) PJM Interconnection, the grid operator across 13 eastern states and the District of Columbia, said on April 10 it wants to line up 15 gigawatts of new power in an emergency process built around expected data center demand. (bloomberg.com) Wood Mackenzie told Bloomberg in March that United States data center development had slowed as the grid hit limits, and developers added about 25 gigawatts of projects in the fourth quarter of 2025, roughly half the third-quarter pace. (bloomberg.com) That is changing the capital stack. Blackstone said on April 10 that it had publicly filed for an initial public offering for Blackstone Digital Infrastructure Trust, a vehicle focused on stabilized, newly constructed data centers. (blackstone.com) Bloomberg reported the new Blackstone vehicle plans to buy already built and leased properties, a sign that investors are paying for deployment certainty rather than waiting on greenfield construction. (bloomberg.com) Private lenders are moving in too. CNBC reported on April 6 that hyperscalers are tapping private credit and debt markets to finance artificial intelligence data center build-outs as insurance, construction and technology risks rise with project size. (cnbc.com) The frictions are physical and local. Bloomberg reported on April 1 that shortages of transformers, switchgear and batteries are delaying projects, and on February 25 it reported that new United States data center construction fell for the first time since 2020 amid permitting, zoning and power-procurement delays. (bloomberg.com, bloomberg.com) The result is an artificial intelligence race that looks more like utility planning and industrial development than software scaling. The winners are increasingly the groups that can secure megawatts, interconnection queues and shovel-ready sites before they secure the next model release. (bloomberg.com, blackstone.com)
