Compute
38 predictions spanning 2026-2035
2026
2027
- A major AI company controls 50% of world AI compute
- Largest training runs exceed $1B
- Annual AI investment reaches $1 trillion
- Frontier training runs need five gigawatts
- U.S. builds fifty AI-dedicated gigawatts
- Millions of AI instances run at 10x-100x human speed
- Effective compute rises five orders of magnitude
2028
2029
2030
- AI uses 7.4% of US electricity
- Chip supply supports 100,000x GPT-4 compute
- Energy bottlenecks do not stop AI scaling
- Frontier models train with 3,000x more compute
- Largest training runs need 4-16 GW
- Largest-model training compute grows 125-fold
- Models need 10-100x less compute
- Data center electricity use reaches 945 TWh
- Data centers drive 20% of advanced-economy demand growth
- Small modular reactors come online for data centers
- U.S. data centers outconsume heavy industry
- Distributed AI training spreads across sites
- Leading AI supercomputer costs $200B
- Leading AI supercomputer reaches 2e22 FLOP/s
- Leading AI supercomputer requires 9 GW
- Leading AI supercomputer uses 2 million chips
- Power becomes main AI-supercomputer bottleneck
- Trillion-dollar cluster reaches 100 GW