The Supply and Demand of AI Tokens | Dylan Patel Interview

Summary

Dylan Patel discusses the explosive demand for AI tokens, his firm's $7M annual token spend, and the supply chain bottlenecks in memory, TSMC, and equipment. He predicts DRAM prices will double or triple, and that semiconductor equipment makers will benefit from surging capex. The conversation covers model progress, the risk of a permanent underclass, and expected public backlash against AI.

• SemiAnalysis’s token spend hit $7M/year, 25% of salary costs.
• Anthropic’s revenue surged from $9B to ~$40B ARR with >70% gross margins.
• Mythos model is a huge step up, but Anthropic hoards it due to compute constraints.
• DRAM supply growth is slow; prices expected to double/triple.
• TSMC capex may reach $100B by 2028, boosting equipment makers.
• CPUs are sold out due to reinforcement learning and inference workloads.
• Dylan predicts large-scale anti-AI protests within three months.
• The value generated by tokens is hard to measure (phantom GDP).