Overview
Potter (Institute for Progress, June 2024) — First in IFP’s “Compute in America” series examining challenges of accelerating the American AI data center buildout. Core argument: AI requires massive physical infrastructure, power is the binding constraint, and losing this race has national security implications.
Scale of the Challenge
- Training GPT-4 required ~21 billion petaFLOP
- On an iPhone 12: 60,000+ years to train GPT-4
- Amazon plans $150B on data centers over 15 years
- Meta plans $37B on AI infrastructure in 2024 alone
- Hyperscalers have enough planned capacity to double their existing footprint
Power as Binding Constraint
- 9 of top 10 US utilities name data centers as their main customer growth source
- Availability and price of power are the top two factors driving data center site selection
- Northern Virginia alone: 2,499 MW existing + 1,237 MW under construction
Data Center Fundamentals
A large modern data center contains tens of thousands of rack-mounted servers. The evolution: from IBM System/360 in 1964 (200-page physical planning manual) through Google’s first 30-server cage to today’s million-square-foot purpose-built facilities.
Key components: IT equipment (servers in racks), power infrastructure (redundant feeds, UPS, generators), cooling systems (the major engineering challenge), and networking.
Strategic Implications
- Advanced AI could massively accelerate scientific progress and economic growth
- Powerful AI systems are critical to national security (offensive and defensive)
- American leadership in clean, firm power should be leveraged for AI infrastructure
- Another transformative technology invented in America could be lost to foreign competitors
Related
- building-modern-data-center — Lowe et al.’s design principles for traditional data center architecture
- homelab-file-management-principles — AP’s practical infrastructure at personal scale