IndexAI and compute
AI and compute
“AI and Compute” is a 2018 OpenAI essay. It argues that the compute in the largest training runs doubled about every 3.4 months in the period they measured, much faster than Moore’s law.
Updated 2026-09-27
Read the essay at openai.com/index/ai-and-compute. The 3.4-month figure is their historical measurement, from the perceptron era through the results they plotted in 2018. It is not a promise that the same slope still holds, and it is not a GPU shopping list.
What the post is actually for
- A chart of compute used by notable training runs
- A comparison against the older 2-year doubling people expected from chips
- The policy point they drew at the time: more compute was a straight path to stronger results, so it needed attention
Modern training budgets are in later posts and in product prices, not in that chart. GPT-5.6 price cuts and Sora’s shutdown are both compute stories from 2026, and neither updates the 2018 slope.
Questions
Who wrote AI and Compute?
OpenAI published it in 2018. The byline is on the essay. Cite that page, not a paraphrase, if you are using the 3.4-month number.
Does the doubling still happen every 3.4 months?
The essay does not claim that for 2026. It reports the era it measured.
Related notes
- GPT-6 and GPT-5.6
People search GPT-6 and land on OpenAI’s GPT-5.6 post. GPT-5.6 ships as Luna, Terra, and Sol. There is no separate GPT-6 product page.
- Sora 2 price and shutdown
Sora’s app closed on 26 April 2026. The API closed on 24 September 2026. There is no Sora 3. Listed API rates were per second of video.
- OpenAI API and API keys
Create an OpenAI API key at platform.openai.com. Usage is billed apart from ChatGPT Plus. new-api is a separate open-source gateway, not the official API.