Dylan Patel – Two labs will soon control most of the world's workforce

Summary

Dylan Patel lays out the economics of AI compute, arguing OpenAI and Anthropic will centralize most usable FLOPs because they monetize compute best. He discusses how massive AI capex and debt issuance could push interest rates higher, stress sovereign borrowers, and derate non-AI equities. The conversation also covers supply-chain bottlenecks, China's compute catch-up, and regulation as the main brakes on takeoff.

• AI infrastructure capex is growing from roughly $1T in 2025 toward $2T+ by 2028 and $11T through 2029.
• OpenAI and Anthropic are taking 40-50% of incremental compute next year and could control most usable FLOPs by 2028.
• EUV tools, memory, and compute capacity are supply bottlenecks with pricing power.
• AI debt issuance could raise rates and trigger sovereign defaults in fragile countries like Pakistan and Nigeria.
• Higher discount rates would crush long-duration non-AI equities and penalize banks.
• China's domestic semiconductor capacity may inflect up in 2028 despite export controls.
• Regulatory limits on model release could slow AI deployment and centralization.