The Supply and Demand of AI Tokens | Dylan Patel Interview

Key points

• Personal data point: Semianalysis (Patel's firm) Claude code spend went from ~$0 to $7m run-rate over 4 months. Salary expense is $25m. AI is now 28% of payroll. 'If this trajectory continues we'll spend more than 100% by year-end.'
• Concrete leverage examples: (1) one analyst built a GPU-accelerated chip-reverse-engineering app overlaying material composition on SEM scans (replaces an entire Intel team's job). (2) Malcolm at a major bank built a 2000-task BLS automation evaluation solo — would have taken 200 economists a year. He coined **'phantom GDP'**: when cost falls faster than output rises, real GDP shrinks even as productivity explodes.
• Energy data set built solo in 3 weeks: every US power plant + every transmission line above a voltage threshold + demand-source mapping + dashboard. Spend: $6k/day of Claude tokens. Customers ask 'how long did this take you?' — the incumbent vendor has 100 people working on the equivalent for a decade.
• **Anthropic margin trajectory:** start of year 30s% gross margin (per leaked round docs). Now floor of ≈72%. Possibly higher because some incremental compute went to research, not inference. Revenue $9B → $40-45B and rising.
• Why margins exploded: demand so high that Anthropic could throttle usage limits, rate limits, free tiers — and customers still pay. 'The reality is tokens are super, super in demand.'
• The new arbitrage: not buying compute. It's directing the tokens you already access toward high-value tasks. 'A year from now the business is arbitraging tokens. Three years from now the model will know how to point them.'
• Mythos: 'the biggest step up in model capabilities in two years.' Anthropic priced it at 5-10x normal token cost, then deliberately held it back from public release because of cyber risk. Opus 4.7 is the 'preferentially worse at cyber' release version. Model card says so explicitly.
• Patel's begging-on-knees moment: he and Leopold (Aschenbrenner?) literally begged an Anthropic cofounder for Mythos access. The cofounder denied it existed.
• Practical consumer advice from Patel: 'If you have capital, get an enterprise Anthropic contract — pay per token, not subscription. Otherwise you'll be rate-limited at exactly the wrong moments.'
• Cost-to-capability collapse: hitting GPT-4-class capability now costs ~1/600 what it did via models like DeepSeek. But no one cares — everyone wants the frontier because the frontier creates the economically valuable outputs. The cheap commodity-tier models go into background tasks.
• Existential implication for incumbents: 'If I don't move fast enough, I'll be commoditised. Energy-data services is a $900m market with 100-person incumbents. We just demonstrated parity in 3 weeks. Everyone needs to ask: who can commoditise me, and how fast?'

Notable quotes

Themes

• AI as 28%-and-rising of payroll
• Anthropic's 30%-to-72% margin step-change
• Phantom GDP and the token-direction arbitrage
• Mythos as the biggest capability step in two years