06-reference

stratechery agents over bubbles

Agents Over Bubbles

Thompson’s definitive statement that AI infrastructure spending is not a bubble, driven by three LLM paradigm shifts: ChatGPT (basic inference), o1 (reasoning with exponential token increase), and Opus 4.5 (functional agents with compounding compute demands). Each paradigm addressed prior flaws: reasoning reduced hallucinations; agents added verification loops and removed the need for humans to manage models directly.

The critical insight for RDCO: agents reduce the number of humans needed for AI to have massive economic impact. A small number of people with agency can deploy multiple agents, each making multiple reasoning calls. This means demand does not require mass consumer adoption to be enormous.

On enterprise economics, Thompson argues companies will use AI not just for cost savings but to replace coordination-heavy human infrastructure with tireless agents. Forward-looking companies will cut deeper, knowing smaller AI-native competitors will otherwise outpace them structurally.

On the AI value chain, Thompson makes his strongest case yet that model+harness integration (Anthropic and OpenAI) is the profit-capturing layer, not infrastructure or applications. Microsoft’s Copilot Cowork proves this: despite wanting model-agnostic products, Microsoft had to use Anthropic’s integrated offering because the harness cannot be separated from the model. This undermines the “models are commodities” thesis and challenges Apple’s strategy of licensing Gemini and building Siri as a separate harness.

RDCO note: This is the single most important Stratechery piece for our thesis. The model+harness integration argument directly validates our architecture (Claude Code + custom orchestration). The enterprise agent adoption curve Thompson describes is exactly the market we serve. Cross-ref: Jaya Gupta moat thesis (06-reference/2026-04-10-jaya-gupta-anthropic-moat.md).