Will a U.S. or non-U.S. entity train an artificial intelligence model at least 10^27 computational operations (FLOPs) before 1 June 2027?

Started Sep 17, 2025 02:00PM UTC
Closing Jun 01, 2027 04:00AM UTC
See more details
Topics
Seasons

Currently, most the largest publicly known AI models are estimated to have been trained using over 1025 floating-point operations of compute (FLOPs), with some recent models thought to have crossed the 1026 threshold. The exact amount of training compute used to train most frontier models is often proprietary, but industry analyses suggest leading models have been trained at this level since at least OpenAI’s ChatGPT-4 release in March 2023 (Epoch AI). According to Epoch AI at the time of question launch, the largest model to date, Grok 4, is estimated to have been trained on 5 x 1026 FLOPs (written as 5e+26).

In 2018, OpenAI estimated that the amount of compute used to train state-of-the-art models doubled every 3.4 months since 2012, growing over 300,000 times in that period (OpenAI, MIT Technology Review). Researchers in 2022, analyzing a broader set of models found the doubling rate to be about every 6 months from 2012 to 2022, with newer large-scale models demanding more resources and slowing the rate of compute scaling (LessWrong, arXiv). As compute requirements grow, organizations are increasingly running into energy, hardware, and financial constraints (Epoch AI, SemiAnalysis).  


Resolution Criteria: 

This question will resolve “Yes” if a model has been trained using at least 1027 FLOPs (which can also be written as 1e+27 in scientific notation). This question will be resolved using direct statements by the developers on the amount of training compute used to train a model or estimates of training compute from EpochAI. Developer statements will take precedence over third-party estimates. In the event that there are no developer statements and EpochAI is no longer providing estimates of training compute, the RFI team will consult with experts and/or estimates from other AI research organizations to make a determination.

EpochAI's Large-Scale AI Models dataset tracks models trained with compute over 1023 FLOPs. The models trained by U.S. or non-U.S. organizations on the largest number of computational operations can be found by setting the following conditions for the table:
  • Sort by “Training compute (FLOP)” from “9 -> 1”
  • Set filter so that it shows records where 
    • “Country (from Organization)” “is exactly” “United States of America” to see models built by U.S. entities, or
    • “Country (from Organization)” “has none of” “United States of America” to see models built by non-U.S. entities
To be considered for resolution, a model’s “Publication date” must be “on or before” “2027-05-31”. 

Notes:
  • Since multiple answers can be correct, the probabilities for “U.S. Entities” and “Non-U.S. Entities” do not need to sum to 100%.
  • In the event that a model is released without information on the training compute, RFI administrators will wait for EpochAI’s estimate before making a determination. If the model qualifies for resolution, the question would then be resolved as of the model’s publication date and forecasts after that date will not be scored. 
  • The AI model’s domain (e.g., language, image generation, biology, etc.) will not affect resolution. 
  • EpochAI’s confidence ratings (“Confident,” “Likely,” “Speculative,” “Unknown”) will not affect resolution. 
  • If training compute information or estimates become available after question opening showing that a model published before question opening used at least 1027 FLOPs for training, this question will be voided.

Files
Tip: Mention someone by typing @username
Cancel