- Critics of AI spending are judging models trained on outdated hardware, ignoring the upcoming massive wave of compute.
- Global AI chip capacity is set to increase 6x between 2024 and 2026, based on TSMC CoWoS wafer allocations.
- Nvidia leads AI silicon supply, with Broadcom (Google TPUs) and AMD also significantly increasing production.
- There's a substantial lag between chip production and actual deployment due to liquid cooling issues and power capacity constraints.
- Training and inference delays mean current models reflect compute from about 12 months ago, not the latest infrastructure.
- Advanced models like Opus 4.5 and Gemini 3 show step-change improvements in software engineering and UI design, likely trained on older compute.
- Future compute (100+ exaFLOPs in 2025, 220+ in 2026) has not yet completed training runs, suggesting even greater advancements ahead.
- By 2030, global AI compute could reach zettascale levels, potentially accelerating AI capabilities dramatically.