AI Trade — Capex Sold, Apps Bought

On Thursday the Financial Times reported that OpenAI told investors its annualised revenue was about $50 billion at the end of September — not the roughly $70 billion that had been circulating. Much of the gap is accounting: OpenAI counts only the revenue it keeps and leaves out sales through cloud partners, while Anthropic includes them. However one explains it, it matters. The whole AI build-out is priced on the belief that the buyers at the end of the chain earn enough to pay for the chips, the data centres and the power — and the biggest buyer turned out to earn less than assumed.

The market answered precisely. The chip index fell 3.4%, Broadcom 4.4%, Micron 4.8%, Oracle 5.5%, the data-centre builders CoreWeave 7.8% and Nebius 7.4%, the power suppliers Vicor 7.2% and Vistra 6.4%. The software that uses AI went the other way: Atlassian +4.0%, Adobe +3.6%, Snowflake +3.2%, the cloud-software fund WCLD +1.5%. That pattern is the thesis this diary has followed since spring.

The thesis: capex first, then opex, then applications

In February our report on agentic infrastructure software drew the line: hardware is a one-time, capex-heavy purchase with cyclical risk — a chip is bought once and depreciates — while the software that runs agents is operating spend, usage-based, and compounds with adoption. On 30 June the Daily Pulse The Opex Handover called the turn: the AI trade was not running out of air, it was relocating from the capex layer to the layer that gets paid when the chips are actually used. The wikifolio AI Cycle 2030 is built on the same three overlapping phases: AI capex and physical bottlenecks, AI opex revenues, AI applications.

We track the chain with three of our own equal-weight indices: Rubin Build-Out for the builders, Agentic Infrastructure for the operating layer, Agentic Winners 40 for the companies that use AI in their products. The AI Handoff Board measures the money moving between them every day.