Tracking AI Gigawatts, Revenue, Expenses and Deep Dive into Productive Agents

admin
By
3 Min Read

SemiAnalysis says Anthropic and OpenAI each ended 2025 with nearly 2 GW of AI conpute and each should end 2026 near 5 GW. Epoch AI data center tracking identifies named location with about 1.45 GW for Anthropic and 421 MW for OpenAI with confidence. Add every announced 2026 project and the end of 2026 number are 4.4 GW short across the two. The shortage is likely AI cloud purchases from Microsoft, Amazon and Google.

I also have a deep dive on ai agents and token economics. A gigawatt of GB200 makes 75–85 quadrillion tokens a year at frontier configuration and about 17 quadrillion as a real mixed fleet. Twenty always-on agents would burn 2.3 trillion tokens a year per person. The top 1% of AI spenders are at those levels of API token usage to use coding agents.

Highlights

* Anthropic’s Q2 expense estimates implies about 2.7 GW (they have the higher prices SpaceX rental for some of their data center) running on average in 2026.

* OpenAI’s spend implies about 2.4 GW.

* Realized industry revenue is about $7.0bn per GW-year against $7.9bn to own and operate one. Revenue per gigawatt roughly equals cost per gigawatt. It is far less than the $100bn per GW per year for frontier AI from Anthropic and OpenAI.

* Codex reached 20 million users, but not 20 million developers. Three releases landed on one day and the metric changed windows.

* AI spending at different levels (source Ramp – top 1%, top 10% and median spending). Top 1% is currently $7,400 a month, which is $88,800 a year. Added a forecast for the next three years where tokens continuing getting cheaper but we still want to use more AI.

Screenshot

Nextbigfuture.substack.com has details and subscribers can get two detailed reports.

Share This Article
Leave a Comment

Leave a Reply