A clear-eyed inventory of the constraints that define the frontier of AI infrastructure today. These are not opinions — they are physics, supply chains and policy made visible.
Air-cooled facilities top out around 30kW/rack. Frontier AI training nodes are now specified at 120–250kW. The gap is not incremental — it requires a different physical paradigm.
Tightly coupled training treats the entire data center as one machine. East-west traffic dominates, and any hop count above three measurably degrades throughput on long-context training runs.
AI workloads create gigawatt-scale, second-resolution load steps. Grid operators were not designed for this. Without on-site buffering, utility infrastructure becomes a hard ceiling on training scale.
Conventional data center builds run 24–36 months. The half-life of a model architecture is currently nine months. Modular and pre-engineered approaches are no longer optional — they are the only viable schedule.
Transformers, switchgear, optical transceivers and high-density CDU systems all face 12–24 month lead times. Strategic procurement is now a core infrastructure competency.
Where compute lives determines who can audit it, regulate it, or seize it. National AI strategies increasingly treat compute infrastructure as critical defense infrastructure.
A small number of engagements per quarter. We work with sovereign funds, frontier labs and hyperscale operators on infrastructure that lasts.