Dylan Patel: Two Labs Will Soon Control Most of the World's Compute
Dwarkesh Patel · youtube · 2026-08-25
Dwarkesh Patel interviews Dylan Patel on the future economics of AI labs. Key takeaways:
- Compute Monopoly: As Recursive Self-Improvement (RSI) nears, training costs will eclipse inference. Anthropic and OpenAI are on track to control most usable global FLOPs because their superior monetization allows them to outbid others for compute.
- Sovereign Debt Risk: The projected $10T+ in AI capex by the end of the decade could trigger a debt crisis. Hyperscaler debt may drive up interest rates, bankrupting non-AI exposed countries and crashing non-AI equities.
- Centralization Forces: The industry trends toward centralization due to economies of scale in training, compute scarcity, and eventual continuous learning/RSI. It remains unclear what can effectively counter these forces.
[Full transcript and podcast links in source]
Related event: Dylan Patel: Two Labs Will Dominate Global Compute and Labor(2 posts)→
More from AGI Musings
- DeepMind Philosopher Discusses Whether Chatbots Are Conscious — dioscuri · 2026-08-26
- Why the public hates datacenters: AI must cure diseases to change minds — adityaag · 2026-08-26
- Vibe coded app slop might end spending on apps — StewartalsopIII · 2026-08-26
- Experiment: Where does adaptation happen in human-LLM interaction? — mb3rtheflame · 2026-08-26
- Science Corporation Focuses BCI on Restoring Vision and Hearing — No Priors · 2026-08-26
- Assessing AI risk like tasting wine is a dangerous intuitive approach — AndyMasley · 2026-08-26