Diary / field note
The leaked OpenAI numbers are a brutal frontier-economics lesson
Frontier AI is not just a capability race. It is a capital-allocation race where revenue can look enormous and still be dwarfed by compute, training and infrastructure.
The leaked OpenAI numbers are the bit of the AI story people keep trying to avoid.
Reported documents say OpenAI booked about $13B in 2025 revenue and spent about $34B in the same year. One reported line item alone: roughly $10.59B paid to Microsoft for research and development expenses, most likely infrastructure.
These are not official public OpenAI filings, so treat them as reported leaked financials rather than gospel.
But the lesson is still useful.
Frontier AI can have enormous demand and still look economically violent.
Revenue is not the same as durability when the product depends on:
- frontier training runs
- inference at massive scale
- Microsoft-class infrastructure
- constant model replacement
- talent markets priced like professional sport
That is why the local and specialist model story matters.
Not because everyone needs to run a toy model on a laptop.
Because the economics of “send everything to the biggest model forever” may not survive contact with the spreadsheet.
The next serious AI stack will be routed. Frontier models where they are needed. Smaller models where they are enough. Local models where privacy or cost matters. Specialist models where the job is narrow and repeated.
Capability gets the headlines.
Unit economics decides what stays.
Source: Where’s Your Ed At and Quartz