Walk through a modern data hall, and the sound has changed. Less white‑noise whoosh, more layered hum—a denser, hotter orchestra of GPUs pulling power like a small town while chilled loops whisper through manifolds that weren’t in last decade’s blueprints. The AI boom isn’t just another upgrade cycle. It’s a buildout so large it needs its own vocabulary: gigawatts as milestones, liquid cooling as default, campuses measured in square miles, and procurement plans that look suspiciously like national infrastructure policy.
Capex at a new altitude
The money is staggering and, importantly, dated. Goldman Sachs pegs global data‑center power demand rising 165% by 2030 versus 2023, with US construction spend tripling in three years and record‑high occupancy even as capacity hits the grid. Deloitte’s power-and-utilities survey reads like a stress test for the grid: AI data‑center load could grow from 4 GW in 2024 to 123 GW by 2035 in the US alone, pushing multi‑gigawatt campuses and seven‑year interconnection queues into the planning baseline. Hyperscaler and utility capex is counted in trillions over the next few years, with hyperscaler outlays expected to hit the trillion‑dollar threshold in roughly three years as inference, not just training, soaks up capacity.
Ten‑gigawatt projects and four million GPUs
The clearest tell of a new class of build is how vendors describe it. Nvidia and OpenAI just put numbers to the scale: a plan for up to 10 GW of AI data centers—roughly four to five million GPUs—funded through an investment that could reach $100 billion in tranches as facilities come online from 2026 onward. Reuters’ shorthand is blunter: about $10 billion per gigawatt, with ~$35 billion per GW going to chips and systems; the rest is power, land, buildings, and everything that keeps silicon cold and honest. It’s not a one‑off either; reports suggest the cash largely loops back into leased Nvidia systems, a vertical stack where capex, supply, and utilization become two sides of the same balance sheet.
Power, cooling, and the laws of physics
All of this runs into three hard walls: electrons, heat, and lead times. Deloitte tallies data‑center power use around 2% of global electricity this year (536 TWh) and reminds that 78% of a typical facility’s draw is split between compute and cooling, with liquid solutions moving from curiosity to requirement as rack densities spike. Brightlio’s market read is equally pragmatic: liquid cooling adoption accelerates across AI designs, not as a brag but as a way to keep hardware alive and PUE defensible under loads that melt yesterday’s airflow assumptions. Even with better thermals, grid interconnection has become the long pole, forcing developers toward colocated generation, PPAs stitched to new load, and “single interconnect” campuses that inject surplus back to the grid at peaks.
The supply chain sprawls
This isn’t just GPUs. It’s HBM, advanced packaging, switch silicon, power electronics, transformers, chillers, and the messy middle of factory tooling. TSMC’s latest brief hits the theme: energy‑efficient logic and packaging are becoming existential design questions as AI stretches from cloud to edge. Chip industry roundups are full of moves that would have been footnotes before—onsemi buying PowerTech rights to feed AI data‑center demand, for instance—because the constraint no longer lives in one aisle.
Geography is shifting, the edge is real
The gravity hasn’t left Northern Virginia or Dublin, but maps are redrawing themselves. New US states join the race as 24/7 AI loads collide with local grids, while the edge thickens to cut latency for inference. Weekly deal sheets now read like travel itineraries: EdgeConneX and Lambda planning dual‑city, 30‑plus‑MW AI sites in Chicago and Atlanta; Middle East mega‑projects pairing exotic cooling with greenfield cities; multi‑gigawatt “districts” scoped before zoning ink dries.
Business models are bending to fit
As build scales past the comfort zone of traditional financing, deals look more like hybrids: infrastructure REITs meet cloud opex; long‑term PPAs sit next to chip‑lease structures; hyperscalers experiment with colocating load and generation to skip transmission waits and stabilize local grids. Bain’s math is bleakly simple: compute demand is outpacing Moore’s law more than 2x, implying 100 GW of new US load by 2030 unless efficiency miracles land on time. Investors, meanwhile, eye “AI infra” as its own asset class—chips, servers, fabs, grid gear, factories—with diversified exposure beyond one vendor’s quarterly.
The floor feels
Stand in a live hall and it’s easy to forget the headlines. The day is gauges and alarms: delta‑T tight, flow steady, harmonics within tolerance. A PPA milestone goes gree,n and someone in legal actually smiles. A substation contractor texts a photo of a transformer finally leaving the yard. In another tab, a scheduler shifts an inference cluster to ride cheap wind, then back again before the evening peak. The drama isn’t a product launch; it’s uptime.
AI’s expansion has crossed a line where the interesting work happens below the marketing layer: power studies, cooling diagrams, interconnect drawings, component purchase orders that read like novels. The winners will be the ones who treat this less like a sprint for square footage and more like a systems problem—where electrons, heat, silicon, and capital learn to keep time together. From here on out, “infrastructure” isn’t backdrop. It’s the story.
