Skip to content
Longterm Wiki
Back

Credibility Rating

4/5
High(4)

High quality. Established institution or organization with editorial oversight and accountability.

Rating inherited from publication venue: Epoch AI

Epoch AI empirical analysis useful for understanding the economics of frontier AI development and informing governance discussions about compute thresholds and access concentration.

Metadata

Importance: 62/100blog postanalysis

Summary

This Epoch AI analysis tracks historical trends in the monetary cost of training machine learning systems, examining how dollar costs have evolved alongside compute scaling. It provides empirical data on training cost trajectories to inform forecasts about future AI development economics and accessibility.

Key Points

  • Training costs for frontier ML models have grown dramatically over time, with top models now requiring billions of dollars in compute expenditure.
  • Despite hardware efficiency improvements reducing cost-per-FLOP, total training costs have risen due to exponentially increasing model scale.
  • Cost trends have implications for who can develop frontier AI systems, concentrating capabilities among well-resourced actors.
  • Historical data enables forecasting of future training costs, relevant for governance and safety planning timelines.
  • Declining per-unit compute costs mean capable models become more accessible over time, raising proliferation concerns.

Cited by 1 page

PageTypeQuality
AI Proliferation Risk ModelAnalysis65.0

Cached Content Preview

HTTP 200Fetched Apr 9, 202698 KB
Trends in the Dollar Training Cost of Machine Learning Systems | Epoch AI
 
 

 
 

 

 

 

 

 

 

 

 
 
 
 

 

 

 

 
 
 

 

 

 

 
 

 
 
 
 
 

 

 

 

 

 
 

 

 

 

 
 
 

 

 
 
 
 

 Oct
 NOV
 Dec
 

 
 

 
 08
 
 

 
 

 2023
 2024
 2025
 

 
 
 

 

 

 
 
success

 
fail

 
 
 
 
 
 
 
 
 
 
 

 

 
 
 
 
 
 
 
 
 

 

 About this capture
 

 

 

 

 

 

 
COLLECTED BY

 

 

 
 Organization: National Archives and Records Administration
 

 

 National Archives and Records Administration crawl performed by Internet Archive. This data is currently not publicly accessible.
 

 

 

 
 
Collection: NARA 118th Congressional Crawl

 

 

 This crawl of online resources of the 118th U.S. Congress was performed by Internet Archive on behalf of the United States National Archives & Records Administration (NARA).
 

 

 

 

 

 
TIMESTAMPS

 

 

 

 

 

 

The Wayback Machine - http://web.archive.org/web/20241108001435/https://epochai.org/blog/trends-in-the-dollar-training-cost-of-machine-learning-systems

 

 

 

 

 
 

 
 
 

 
 

 

 
 

 

 

 
 
 Articles
 

 
 
 
 Research
 

 
 
 
 Trends
 

 
 
 
 Data
 

 
 
 
 Team
 

 
 
 
 Careers
 

 
 
 
 Support us
 

 
 
 
 

 
 
 
 

 

 
 
 
 

 

 
 

 

 

 

 Menu
 

 
 

 
 

 

 

 Close
 

 
 

 
 

 

 

 

 

 

 

 

 
 Articles
 
 Research
 
 Trends
 
 Data
 
 Team
 
 Careers
 
 Support us
 
 
 Search
 

 
 
 
 

 

 
 
 
 

 
 

 

 
 
 
 
 

 

 

 

 

 
 Search epochai.org
 

 
 Search
 

 

 

 

 Enter a query to search for results
 

 

 

 

 
 

 

 Placeholder

 

 

 

 

 

 
 

 

 

 

 Articles
 
 
 

 
 
 
 Trends in the Dollar Training Cost of Machine Learning Systems
 

 

 
report

 
Trends in the Dollar Training Cost of Machine Learning Systems

 
 
 
I combine training compute and GPU price-performance data to estimate the cost of compute in US dollars for the final training run of 124 machine learning systems published between 2009 and 2022, and find that the cost has grown by approximately 0.5 orders of magnitude per year.

 

 

 

 

 
 
 
 

 

 

 

 

 

 

 
Published

 
Jan 31, 2023

 

 

 
 
 
 
 
 

 

 
Authors

 
 
 Ben Cottier
 
 
 

 
 

 
Resources

 

 

 
 

 

 

 
 
 Dataset
 
 

 
 

 

 

 
 
 Source Code
 
 

 
 
 Cite
 
 

 

 

 

 

 

 

 
 
 

 

 

 

 Contents 
 
 

 
 

 

 

 

 

 

 

Summary

Why study dollar training costs?

Method

Background on methods to estimate the dollar cost of training compute

Estimating training cost from training compute and GPU price-performance

Method 1: Using the overall GPU price-performance trend

Method 2: Using the price-performance of actual hardware used to train ML systems

Dataset

Code

Large-scale systems

Results

Method 1: Using the overall GPU price-performance trend for all ML systems (n=124)

Growth rate of training cost for all ML systems: 0.51 OOMs/year

Growth rate of training cost for large-scale ML systems: 0.2 OOMs/year

Method 2: Using the price-performance

... (truncated, 98 KB total)
Resource ID: dff8fae99b47e61d | Stable ID: YTQ5YmRiYz