Dylan Patel describes explosive, near-unbounded demand for frontier AI tokens, driven by users’ strong preference for the latest models—exemplified by SemiAnalysis’s AI spend surging from tens of thousands to $7 million annually, now exceeding 25% of payroll. Claude Mythos represents the largest capability leap in two years, operating at an L6 software engineer level while improving token efficiency. Demand is accelerating further with robotics poised as the next major token consumption wave. On the supply side, severe bottlenecks span memory (DRAM), logic chips, wafer fab equipment, and even CPUs—now sold out industry-wide—amid TSMC’s record $100B capex plan straining downstream suppliers. While scaling laws still hold, compute scarcity is inflating margins for top labs and forcing tough trade-offs: firms must deploy more tokens to capture value or risk permanent competitive disadvantage. Tokenomics remains elusive in practice—real usage consistently outpaces estimates—and rising public backlash warns AI leaders to shift messaging toward tangible benefits and away from hype to avoid imminent protests.