Dylan Patel - The Infinite Demand for Tokens, Claude Mythos, and Supply Constraints - [Invest Like the Best, EP.468]

What It Covers

Dylan Patel of Semianalysis details how AI token demand is growing faster than infrastructure can supply it, using Semianalysis's own spending trajectory from tens of thousands to $7M annually as a case study, while mapping semiconductor bottlenecks in memory, logic, and fab equipment that constrain scaling through 2028.

Key Questions Answered

• •Token spend as competitive moat: Enterprise AI contracts with Anthropic now include rate limit increases as a strategic asset. Firms willing to pay per-token rather than subscription avoid usage caps. Semianalysis reached 25% of salary expense on Claude Code alone, with projections suggesting AI spend could exceed total payroll by year-end if current growth continues.
• •Frontier model premium is non-negotiable: Users immediately abandon previous model versions the moment a new frontier model releases, regardless of cost. Anthropic's Mythos is priced 5–10x higher per token than standard models yet demand exceeds supply. Willingness to pay scales with model capability because economic value generated per token grows faster than token cost.
• •DRAM prices will double or triple from current levels: Memory capacity can only grow 20–30% annually, and new fab capacity decisions made now won't produce output until 2027–2028 at earliest. The only mechanism to balance demand against constrained supply is price-driven demand destruction. Investors underestimating this timeline are mispricing memory-exposed positions in the semiconductor supply chain.
• •Implementation cost collapse reorders competitive advantage: When AI reduces execution difficulty to near-zero, the scarce resource shifts entirely to idea selection and capital allocation. One Semianalysis economist, working alone with Claude, replicated work previously requiring a 200-person bank economics team in days, including a novel 2,000-task AI capability benchmark measuring deflationary GDP effects.
• •TSMC CapEx trajectory points toward $100B annually by 2028: Current 2025 CapEx sits at $57–58B. Downstream equipment suppliers like ASML, Lam Research, and Applied Materials face compounding demand as TSMC scales. Copper foil, glass fiber, and laser supply chains are already constrained. Investors should track second and third-tier semiconductor equipment names for supply-driven margin expansion ahead of consensus estimates.