Gigawatt-scale AI compute
Data-center clusters of hundreds of thousands to millions of accelerators with the power, cooling and interconnect to train and run frontier models.
State of the art
Multiple 100k–200k GPU clusters operating in 2025–2026; gigawatt campuses announced.
- Grid power and interconnection queues
- Advanced-node chip and HBM memory supply
Who’s working on it
- NVIDIA ↗Company
Supplies the GB200/GB300 Blackwell racks (72 GPUs per NVLink domain) that every gigawatt-class 2025-2026 cluster is built from
- Stargate (OpenAI, Oracle, SoftBank, Crusoe)Program
$500B US data-center initiative; the Abilene, Texas flagship energized its first buildings in September 2025 and ran ~0.3 GW of a 1.2 GW design by April 2026
- Microsoft Azure Fairwater ↗Company
Announced Fairwater Wisconsin (September 2025) and Fairwater Atlanta (November 2025), linked into one "AI superfactory" with hundreds of thousands of Blackwell GPUs on a flat network
- xAI Colossus 2Company
Memphis site reached roughly 1.8M H100-equivalents at ~946 MW IT power by mid-2026, at an estimated $36B spent of a $58B projected build
- Meta HyperionCompany
Richland Parish, Louisiana supercluster scaled from a ~$10B plan to a 5 GW, $50B+ build announced July 2026
- Ben Cottier ↗ResearcherEpoch AI
Leads the Frontier Data Centers Hub (2025-2026), the main independent accounting of cluster power, cost and chip counts
- TSMCCompany
Sole-source for leading-edge accelerator dies and CoWoS packaging; its advanced-packaging capacity, not design, has been the binding constraint since 2024
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| infrastructure | Firm power generation and transmission for data centers | $$$$ > $1B |
Depends on
A root node: no prerequisites modeled.