The investment thesis
The $228B spent in 2024 wasn't buying GPT-4 inference — that model already ran on 2023 hardware. It funded training clusters for models shipping 2–3 years later, and inference infrastructure for the next generation. The $700B guided for 2026 funds models whose architectures may not be designed yet.
The risk
DeepSeek V3 achieved frontier performance at 10× less compute. If algorithmic efficiency leaps continue, today’s massive clusters could be overbuilt for training. But labs are betting inference will dominate — ~70% of AI compute by 2030. The bet is on deployment scale, not just model size.
From dollar to model: 2.5–4+ years — in ~2-year staircases
Building a datacenter takes 12–24 months. Procuring chips takes 6–12 months. Training takes 6–12 months. Post-training and safety adds 3–6 months. Each year’s capex splits into three bets: inference capacity arriving in ~2 years, training runs for models ~3 years out, and research compute powering breakthroughs 4+ years away. Progress follows a staircase pattern: a big capability jump every ~2 years, then refinement within that generation (GPT-4 → 4 Turbo → 4o, or Claude 3 Opus → 3.5 Sonnet → Opus 4.6).
⚠️ Figures for 2026 onward (capex guidance, model names, parameter counts) are projections and industry estimates, not confirmed facts — labs no longer disclose parameter counts, and spending plans shift quarterly. Treat this as a thesis to discuss, not a datasheet.