The Supply and Demand of AI Tokens | Dylan Patel Interview

Dylan Patel of SemiAnalysis argues Anthropic’s Mythos model represents the biggest capability leap in two years and predicts large-scale anti-AI protests within three months.

• SemiAnalysis AI spend hit $7M/year, over 25% of their $25M salary base, accelerating toward exceeding headcount cost entirely.
• Anthropic gross margins jumped from ~35% (leaked funding docs) to at least 72%, with revenue growing ~$10B/month.
• Mythos (unreleased) benchmarks suggest it performs like an L6 software engineer; Anthropic went from L4 to L6 in two months.
• Mythos is 5–10x the token cost of Opus 4.6 but cheaper per task because it uses far fewer tokens to complete work.
• DRAM prices will double or triple again; incremental memory capacity decisions made now won’t arrive until 2027–2028 at earliest.
• TSMC capex may reach $100B in 2028, creating severe tail-whip shortages across wafer equipment, optics, copper foil, and PCB supply chains.
• CPUs are sold out due to two new demand sources: reinforcement learning environments and deployed AI-generated code running on cloud instances.
• Patel predicts large-scale protests against Anthropic and OpenAI within three months, citing AI’s approval rating below ICE and politicians.

2026-04-23 · Watch on YouTube