SemiAnalysis's Claude Code spend is about $7m a year against roughly $25m in salaries, and could pass 100 percent of salary by year-end, Dylan Patel says.
Anthropic's revenue has grown from about $9bn to an estimated $35 to $40bn annualized, with its revenue adding $10bn a month, which outpaces its compute.
Patel says Mythos may be the biggest capability jump in two years, yet Anthropic is reluctant to release it, while GPU prices rise and memory supply stays tight until 2028.
Token demand is outrunning the AI supply chain
Demand for AI tokens is outrunning supply, and compute is the binding constraint. Anthropic's revenue has grown from $9bn to $35 to $40bn annualized while customers still capture more value than they pay. Every layer from GPUs to memory is sold out, and new memory capacity does not arrive until late 2027 or 2028.
Takeaways
Anthropic demand far exceeds its compute. Anthropic's annualized revenue rose from $9bn to roughly $35 to $40bn and adds $10bn a month. It can cut usage and rate limits because customers capture far more value than they pay, and Patel estimates gross margins of at least 72 percent.
SemiAnalysis shows how fast AI spend scales. Its Claude Code spend runs about $7m a year against roughly $25m of salaries, over 25 percent and possibly past 100 percent by year end. Non-technical staff now build tools, such as an energy analyst's map of every US power plant.
Frontier models cost more while capability gets cheaper. The cost of reaching a given capability tier has fallen 100 to 1000 times, and Patel expects another hundredfold drop in a year. Mythos costs five to ten times more per token yet uses fewer tokens, so most tasks cost less than on Opus 4.6.
Cheap implementation concentrates tokens among fewer companies. Release cadence fell from six months to about two. Anthropic largely hit its L4 engineer goal and Mythos benchmarks like an L6. With ideas cheap, only good ideas justify spend, so tokens pool with firms that can afford the first billions.
The whole chip supply chain is sold out. Memory capacity grows only 20 to 30 percent a year, with real new supply in late 2027 or 2028, so DRAM prices could double or triple. ASML, copper foil and CPUs are sold out, and TSMC raises prices only in single digits.
GPUs last longer and cost more. Useful life is seven or eight years rather than five, and Hopper clusters are re-signing. Nvidia still earns about 75 percent gross margins as the supply chain pays large prepayments.
Patel expects public backlash against AI. He predicts large protests within three months, saying AI is less popular than ICE and politicians. He criticises interviews by Altman and Amodei, including Altman on Tucker Carlson, and wants AI companies to talk about present, uplifting uses.
Patel's predictions
The economy spends $100bn on Opus 4.6-tier models by the end of 2026, up from $40bn now.
Anthropic holds 70-plus percent gross margins on that tier, while OpenAI would charge around 50 percent if it got there first.
Even tier-two labs sell out of tokens.
TSMC spends $100bn on capital expenditure in 2028, against $56bn planned for 2026.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Robotics sees few-shot learning breakthroughs in six to 18 months, and vision-language-action models are not the approach that scales.