Latest
Publications & Commentary
Papers & Reports Newsletter Podcast
Data & Resources
Datasets
Overview Benchmarking Models Frontier Data Centers Hardware Companies Chip Sales Polling
Resources
AI Trends & Statistics Data Insights
Projects
FrontierMath GATE Playground Distributed Training Model Counts
About
About Us Our Team Careers Consultations Our Funding Donate
Contact
Model Performance
AI Companies
Hardware
Training runs
Data Centers

Machine Learning Trends

Our ML Trends dashboard offers curated key numbers, visualizations, and insights that showcase the significant growth and impact of artificial intelligence.

Last updated on Dec 9, 2025

Display growth values in:

The Epoch Capabilities index

Confident
+15.5 ECI per year
Model capabilities have grown faster since the start of 2024.
90% confidence interval: +13 to +18 ECI per year.

Compute stock growth

Likely
199.5 x/year
The total computing power of the stock of NVIDIA chips is growing by .
90% confidence interval: to .

FLOP/s per dollar

Confident
23.4 x/year
Performance per dollar improves by each year.
90% confidence interval: to .

Training compute

Confident
100000.0 x/year
Training compute for frontier language models has been growing by since 2020.
90% confidence interval: to .

Algorithmic progress

Likely
1000.0 %/year
The compute required to achieve a given level of prediction accuracy in language model pre-training is declining at a rate of .
90% confidence interval: to .

Largest AI data center

Likely
500,000 H100e
The Microsoft Fairwater Atlanta data center has a compute capacity of around 500,000 H100-equivalents, supported by 600 megawatts of power capacity and costing $20 billion in capex.
90% confidence interval: 250,000 to 1,000,000.

Model Performance

LLM inference prices are falling rapidly

Likely
1.0e+40 x
The cost to inference an LLM at a fixed level of performance has fallen rapidly, but unevenly across tasks.
90% confidence interval: to .

The Epoch Capabilities index

Confident
+15.5 ECI per year
Model capabilities have grown faster since the start of 2024.
90% confidence interval: +13 to +18 ECI per year.
Database

Benchmarking Hub

Our database of benchmark results, featuring the performance of leading AI models on challenging tasks. It includes results from benchmarks evaluated internally by Epoch AI as well as data collected from external sources. Explore trends across time, by benchmark, or by model.

Read more

SOTA vs. consumer hardware models

Likely
8 months
Frontier AI performance becomes accessible on consumer hardware within a year.
90% confidence interval: 6 to 12 months.

LLM context windows

Confident
1.0e+30 x
LLM context windows have grown by since 2023.
90% confidence interval: to .

AI Companies

Compute stock growth

Likely
199.5 x
The total computing power of the stock of NVIDIA chips is growing by .
90% confidence interval: to .

OpenAI revenue growth

Confident
1584.9 x
OpenAI’s annualized revenue has grown by since 2024.
90% confidence interval: to .
Database

AI Companies Hub

Our database of AI company data, with data on revenue, funding, staff, and compute for many of the key players in frontier AI.

Read more

Frontier inference compute allocation

Likely
30 %
The majority of OpenAI’s compute spending in 2024 was allocated to R&D (training and experiments), rather than inference.

Frontier AI investment

Likely
$97 billion
Frontier AI labs have raised more than $90 billion.

Hardware Trends

FLOP/s per dollar

Confident
23.4 x
Performance per dollar improves by each year.
90% confidence interval: to .

Memory bandwidth

Confident
19.1 x
Memory bandwidth of GPUs has been growing by 28% a year since 2008.
90% confidence interval: to .
Report

Trends in Machine Learning Hardware

FLOP/s performance in 47 ML hardware accelerators doubled every 2.3 years. Switching from FP32 to tensor-FP16 led to a further 10x performance increase. Memory capacity and bandwidth doubled every 4 years.

Read more Dataset

Adoption of Numerical Formats

Confident
3-4 years
Widespread adoption of new numeric formats took 3-4 years in past cycles.

FLOP/s per watt

Confident
21.9 x
GPUs have become more energy efficient per year since 2008.
90% confidence interval: to .

Training runs

Training compute

Confident
100000.0 x
Training compute for frontier language models has been growing by since 2020.
90% confidence interval: to .

Algorithmic progress

Likely
1000.0 %
The compute required to achieve a given level of prediction accuracy in language model pre-training is declining at a rate of .
90% confidence interval: to .

Most compute used in a training run

Plausible
5e26 FLOP
We estimate the total training compute for the final training run of Grok 4 was 5e26 FLOP.
Report

Training compute of frontier AI models grows by 4-5x per year

Our expanded AI model database shows that the compute used to train recent models grew 4-5x yearly from 2010 to May 2024. We find similar growth in frontier models, recent large language models, and models from leading companies.

Read more Dataset

Training cost

Likely
3162.3 x
The cost to train frontier language models has increased by since 2020.
90% confidence interval: to .

Training hardware cost

Likely
398.1 x
The cost to purchase the hardware used to train frontier language models has grown by since 2020.
90% confidence interval: to .

Data Centers

Largest AI data center

Likely
500,000 H100e
The Microsoft Fairwater Atlanta data center has a compute capacity of around 500,000 H100-equivalents, supported by 600 megawatts of power capacity and costing $20 billion in capex.
90% confidence interval: 250,000 to 1,000,000.

Build time

Likely
2.1 years
Based on existing projects and future plans, the median time from a data center starting construction to achieving 1 gigawatt of total facility power is about 2 years.
Range: 1 to 3.6 years.
Database

Frontier Data Centers

Open database of AI data centers using satellite and permit data to show compute, power use, and construction timelines.

Read more

Cost per gigawatt

Likely
$30 billion
The total capital cost to build a data center with 1 gigawatt of total facility power is around $30 billion dollars.
90% confidence interval: 20 to 40 billion.

Country with the most computing power

Likely
United States (75% globally)
As of May 2025, the United States contains about three-quarters of global GPU cluster performance, with China in second place with 15%.

Acknowledgements

We thank Tom Davidson, Lukas Finnveden, Charlie Giattino, Zach Stein-Perlman, Misha Yagudin, Jai Vipra, Patrick Levermore, Carl Shulman, Ben Bucknall and Daniel Kokotajlo for their feedback.

Several people have contributed to the design and maintenance of this dashboard, including Jaime Sevilla, Pablo Villalobos, Anson Ho, Tamay Besiroglu, Ege Erdil, Ben Cottier, Matthew Barnett, David Owen, Robi Rahman, Lennart Heim, Marius Hobbhahn, David Atkinson, Keith Wynroe, Christopher Phenicie, Nicole Maug, Aleksandar Kostovic, Alex Haase, Robert Sandler, Edu Roldan, Andrew Lucas and Yafah Edelman.

Use this work

Epoch AI’s work is free to use, distribute, and reproduce provided the source and authors are credited under the Creative Commons Attribution license.

Cite this work as

Epoch AI (2023), "Key Trends and Figures in Machine Learning". Published online at epoch.ai. Retrieved from: 'https://epoch.ai/trends' [online resource]

BibTeX citation

  @misc{epoch2023aitrends,
    title="Key Trends and Figures in Machine Learning",
    author={{Epoch AI}},
    year=2023,
    url={https://epoch.ai/trends},
    note={Accessed: }
  }

If you spot an error or would like to provide feedback, please reach out at .

Thank you for your feedback!

Your comment will be reviewed. We may not be able to respond to every submission.