Anthropic's $11 Billion Month
The number that anchors this conversation: Anthropic added $11 billion in annual recurring revenue in a single month. Palantir, Snowflake, and Databricks — the three standout SaaS businesses of the last decade — took ten years and tens of thousands of employees to collectively reach a similar figure. The structural reason Anthropic got there on roughly 80% less capital than OpenAI comes down to a lower cost-per-token, achieved through disciplined architecture choices rather than brute-force spending.
The Bottleneck Is No Longer Silicon — It's Power
A single Blackwell rack weighs 3,000 pounds, stands eight feet tall, and consumes 100 kilowatts of electricity. The limiting factor for AI infrastructure has migrated entirely away from chip design and into energy and zoning. The US natural gas price advantage — down 20% domestically while European and Asian prices doubled or tripled — is functioning as a hidden subsidy for American AI infrastructure that few analysts are pricing in. The physical build-out for 2026 to 2028 is already committed; the real capital planning question concerns 2030 to 2032.
TSMC Is Deliberately Preventing a Bubble
The leadership running TSMC views their manufacturing monopoly as a generational legacy to protect, not a revenue opportunity to maximise. If they expanded capacity at the rate Nvidia's Jensen Huang has requested, the market would be flooded with two to three trillion dollars worth of GPUs, guaranteeing a crash. Their deliberate constraint is what's preventing a 2000-style dot-com overbuild. The fragility in this arrangement is specific: if Intel or Samsung break ranks and scale above 30% market share as a second-source fab, TSMC's discipline collapses and the classic bubble-and-crash cycle becomes likely.
Open Source Is Living off Stolen Frontier Tokens
The conversation on open source is pointed. Most of the capability in leading open-source models exists because developers allegedly distilled outputs from proprietary frontier APIs to train their own systems. DeepSeek is cited as having used roughly 150,000 reasoning traces from commercial APIs to build their efficient open-source architecture. This creates a prisoner's dilemma for the frontier labs: if they all closed their APIs simultaneously, open-source distillation would die overnight. But any single lab that defects to capture developer revenue hands its outputs to competitors, accelerating the commoditisation of its own technology.
The Only Way to Survive the Application Layer
Of Jensen Huang's five-layer AI stack — energy, data centres, chips, models, applications — profits are currently accruing to every layer except applications. The one clear exception is coding tools: Cursor, Cognition, and Replit are working because coding is the shortest path to a system that can build anything else. The broader lesson from those who've survived is that incrementalism fails. Companies must pick architectural bets so physically difficult — like Cerebras building chips the size of an entire silicon wafer — that incumbents will refuse to copy them, rather than compete directly with Nvidia on Nvidia's terms.