- AI critics are judging models trained on last-gen hardware, but a 6x wave of compute is already allocated and starting to produce results.
- TSMC's CoWoS allocations show NVIDIA leading with 60% share by 2026, Broadcom at 15%, and AMD growing 15x for their AI chips.
- Total AI silicon supply is estimated to grow from 117,000 wafers in 2023 to 1 million by 2026.
- Napkin math estimates a 6x increase in global AI chip capacity between 2024 and 2026, reaching 122.6 exaFLOPs by 2026.
- Significant delays exist between chip production and deployment, including liquid cooling challenges for NVIDIA's GB200 series and power capacity constraints.
- Training models takes at least 6 months, meaning current models reflect infrastructure from ~12 months prior.
- Not all compute is for training; inference and techniques like agentic reinforcement learning are growing in importance.
- Opus 4.5 and Gemini 3 represent major leaps in AI capabilities, trained on 2024's compute (~36 exaFLOPs).
- The 100+ exaFLOPs coming online in 2025 and 220+ in 2026 haven't yet produced trained models, hinting at even greater future advancements.
- By 2030, AI compute could reach zettaFLOP scale (30x 2026 levels), making the scaling debate even more critical.