AI infrastructure

Core thesis

AI infrastructure spending has scaled to a ~$700-725B combined big-four capex run-rate in 2026, but the constraint long ago shifted from chip design to delivery. Grid connection and on-site generation decide when capacity comes online; HBM and CoWoS packaging cap how many accelerators can be built; and optical interconnect is now the third limit as coppers reach fails at cluster bandwidths. Chips can be ordered; the power, packaging and data-movement to deploy them at gigawatt scale are the active limits. Facilities that arent built cant be filled with NVIDIA Blackwell or Rubin.

State of the art (2026)

State of the art (2026). The AI build-out is running at gigawatt scale - OpenAI's Stargate sites and xAI's Colossus expansion are measured in GW, not megawatts - and the binding limits are no longer GPUs alone but power and the data-centre interconnect. Grid connection and on-site generation gate when capacity comes online (see brief 962); HBM and CoWoS packaging cap how many accelerators can be built; and optical interconnect has become the third constraint as copper's reach fails at the bandwidth the clusters now need (brief 973). The competitive question is who can secure all three - power, packaging and optics - not who has the fastest chip.

Power supply for AI data centres

Electricity is one of the main limits on new AI capacity. Nvidia's GB200 and GB300 NVL72 systems put 72 GPUs in a single liquid-cooled rack, so one data hall needs far more power than older designs. Technology companies have signed nuclear agreements to secure supply. Helion plans to start its Orion fusion plant in 2028 and supply at least 50 MW under its agreement with Microsoft, though no fusion plant has yet delivered power to a grid. Google's agreement with Kairos Power aims for a first small reactor in 2030. We do not have verified figures for power use per rack or per training run that hold across sites.