The AI compute race has a new bottleneck — and it isn’t GPUs.
Nvidia will sell as many Blackwell racks as anyone can pay for. What separates the operators pulling ahead from the ones falling behind is how fast they can turn a piece of land into energized megawatts.
This case study maps four live buildouts that took four different routes to power — and what each reveals about where the industry is headed next:
- xAI’s Colossus campus in Memphis went from groundbreaking to a 100,000-GPU cluster in 122 days, by refusing to wait on the grid.
- Microsoft’s Fairwater network linked Wisconsin and Atlanta over 120,000+ miles of new fiber, betting on distributed scale instead of a single site.
- Stargate’s Abilene campus shows what disciplined capital reallocation looks like when a hyperscaler passes on stranded capacity — and a rival moves in within weeks.
- Amazon’s Project Rainier proves a well-structured utility partnership can match vertical integration for speed.
Beneath all four sits a longer bet: the four largest US hyperscalers have now contracted more than 10 GW of nuclear power combined, hedging today’s buildout against tomorrow’s grid.
Goldman Sachs Research projects US data center power demand nearly doubling by 2027 — with only half of scheduled capacity expected to arrive on time. For CXOs, investors, and infrastructure operators, that gap is no longer a facilities problem. It’s a capital allocation problem.
Read the full case study to see the verified numbers, the five capital-allocation actions we’d recommend to any board underwriting AI infrastructure through 2027–2030, and the charts breaking down exactly how the power math works.
