Data Insights

Epoch AI’s data insights break down complex AI trends into focused, digestible snapshots. Explore topics like training compute, hardware advancements, and AI training costs in a clear and accessible format.

Featured

Global AI computing capacity is doubling every 7 months
Data Insight
Jan. 9, 2026
Global AI computing capacity is doubling every 7 months
AI capabilities progress has sped up
Data Insight
Dec. 23, 2025
AI capabilities progress has sped up
LLM inference prices have fallen rapidly but unequally across tasks
Data Insight
Mar. 12, 2025
LLM inference prices have fallen rapidly but unequally across tasks

Filter

Topic
0 results
Claude users skew towards higher-income households; Meta towards lower-income
Data Insight
Apr. 22, 2026
Claude users skew towards higher-income households; Meta towards lower-income

By Caroline Falkman Olsson and Jaeho Lee

Claude usage rose by over 40% amid increased attention but remains far behind ChatGPT
Data Insight
Apr. 15, 2026
Claude usage rose by over 40% amid increased attention but remains far behind ChatGPT

By Yafah Edelman, Caroline Falkman Olsson, and Jaeho Lee

Five hyperscalers now own over two-thirds of global AI compute
Data Insight
Apr. 14, 2026
Five hyperscalers now own over two-thirds of global AI compute

By Luke Emberson, Josh You, and Venkat Somala

Google controls the most AI computing power, driven by its custom TPUs
Data Insight
Apr. 7, 2026
Google controls the most AI computing power, driven by its custom TPUs

By Luke Emberson, Josh You, and Venkat Somala

Total AI chip memory bandwidth has grown 4.1x per year, now reaching 70 million terabytes per second
Data Insight
Mar. 24, 2026
Total AI chip memory bandwidth has grown 4.1x per year, now reaching 70 million terabytes per second

By Luke Emberson

Advanced packaging and HBM, not logic dies, were the bottlenecks on AI chip production in 2025
Data Insight
Mar. 12, 2026
Advanced packaging and HBM, not logic dies, were the bottlenecks on AI chip production in 2025

By Venkat Somala

Microsoft’s recent $68 billion in physical assets additions were driven by AI-related purchases
Data Insight
Mar. 5, 2026
Microsoft’s recent $68 billion in physical assets additions were driven by AI-related purchases

By Isabel Juniewicz

Hyperscaler capex has quadrupled since GPT-4's release
Data Insight
Feb. 26, 2026
Hyperscaler capex has quadrupled since GPT-4's release

By Isabel Juniewicz

Anthropic could surpass OpenAI in annualized revenue by mid-2026
Data Insight
Feb. 19, 2026
Anthropic could surpass OpenAI in annualized revenue by mid-2026

By Luke Emberson and Yafah Edelman

Compute accounts for the majority of expenses of AI companies
Data Insight
Feb. 4, 2026
Compute accounts for the majority of expenses of AI companies

By Luke Emberson and Yafah Edelman

Benchmark scores are well correlated, even across domains
Data Insight
Jan. 23, 2026
Benchmark scores are well correlated, even across domains

By Luke Emberson and Yafah Edelman

Global AI power capacity is now comparable to peak power usage of New York State
Data Insight
Jan. 16, 2026
Global AI power capacity is now comparable to peak power usage of New York State

By Yafah Edelman, Josh You, Venkat Somala, and Luke Emberson

Global AI computing capacity is doubling every 7 months
Data Insight
Jan. 9, 2026
Global AI computing capacity is doubling every 7 months

By Josh You, Venkat Somala, Yafah Edelman, and Luke Emberson

Chinese AI models have lagged the US frontier by 7 months on average since 2023
Data Insight
Jan. 2, 2026
Chinese AI models have lagged the US frontier by 7 months on average since 2023

By Luke Emberson

AI capabilities progress has sped up
Data Insight
Dec. 23, 2025
AI capabilities progress has sped up

By Yafah Edelman and Jaeho Lee

GPUs account for about 40% of power usage in AI data centers
Data Insight
Dec. 18, 2025
GPUs account for about 40% of power usage in AI data centers

By Luke Emberson and Ben Cottier

NVIDIA’s B200 costs around $6,400 to produce, with memory accounting for half
Data Insight
Dec. 10, 2025
NVIDIA’s B200 costs around $6,400 to produce, with memory accounting for half

By Venkat Somala

Today’s largest data center can do more than 20 GPT-4-scale training runs each month
Data Insight
Dec. 4, 2025
Today’s largest data center can do more than 20 GPT-4-scale training runs each month

By Jaeho Lee

Microsoft’s Fairwater datacenter will use more power than Los Angeles
Data Insight
Nov. 26, 2025
Microsoft’s Fairwater datacenter will use more power than Los Angeles

By Jaeho Lee

The largest AI data center campuses will soon be a fifth the size of Manhattan
Data Insight
Nov. 19, 2025
The largest AI data center campuses will soon be a fifth the size of Manhattan

By Ben Cottier

Build times for gigawatt-scale data centers can be 2 years or less
Data Insight
Nov. 10, 2025
Build times for gigawatt-scale data centers can be 2 years or less

By Venkat Somala and Ben Cottier

Epoch’s Capabilities Index stitches together benchmarks across a wide range of difficulties
Data Insight
Nov. 6, 2025
Epoch’s Capabilities Index stitches together benchmarks across a wide range of difficulties

By Jaeho Lee and Luke Emberson

Open-weight models lag state-of-the-art by around 3 months on average
Data Insight
Oct. 30, 2025
Open-weight models lag state-of-the-art by around 3 months on average

By Luke Emberson

OpenAI's revenue has been growing 3x a year since 2024
Data Insight
Oct. 14, 2025
OpenAI's revenue has been growing 3x a year since 2024

By Venkat Somala

Most of OpenAI’s 2024 compute went to experiments
Data Insight
Oct. 10, 2025
Most of OpenAI’s 2024 compute went to experiments

By Josh You

AI capabilities have steadily improved over the past year
Data Insight
Sep. 30, 2025
AI capabilities have steadily improved over the past year

By Luke Emberson

AI developers accurately report GPQA Diamond scores for recent models
Data Insight
Sep. 19, 2025
AI developers accurately report GPQA Diamond scores for recent models

By Jaeho Lee and Yafah Edelman

What did it take to train Grok 4?
Data Insight
Sep. 12, 2025
What did it take to train Grok 4?

By James Sanders, Luke Emberson, and Yafah Edelman

LLMs have not yet solved the hardest problems on high school math contests
Data Insight
Sep. 3, 2025
LLMs have not yet solved the hardest problems on high school math contests

By Greg Burnham

GPT-5 and GPT-4 were both major leaps in benchmarks from the previous generation
Data Insight
Aug. 29, 2025
GPT-5 and GPT-4 were both major leaps in benchmarks from the previous generation

By Luke Emberson and Josh You

Frontier AI performance becomes accessible on consumer hardware within a year
Data Insight
Aug. 15, 2025
Frontier AI performance becomes accessible on consumer hardware within a year

By Venkat Somala and Luke Emberson

Compute is not a bottleneck for robotic manipulation
Data Insight
Aug. 8, 2025
Compute is not a bottleneck for robotic manipulation

By Ben Cottier, Scott Longwell, James Sanders, David Owen, Yafah Edelman, and Luke Emberson

Training open-weight models is becoming more data intensive
Data Insight
Aug. 1, 2025
Training open-weight models is becoming more data intensive

By Venkat Somala and Yafah Edelman

Frontier training runs will likely stop getting longer by around 2027
Data Insight
Jul. 25, 2025
Frontier training runs will likely stop getting longer by around 2027

By Luke Emberson and Yafah Edelman

LLMs now accept longer inputs, and the best models can use them more effectively
Data Insight
Jun. 25, 2025
LLMs now accept longer inputs, and the best models can use them more effectively

By Greg Burnham and Tom Adamczewski

LLM providers offer a trade-off between accuracy and speed
Data Insight
Jun. 11, 2025
LLM providers offer a trade-off between accuracy and speed

By Greg Burnham and Tom Adamczewski

Acquisition costs of leading AI supercomputers have doubled every 13 months
Data Insight
Jun. 5, 2025
Acquisition costs of leading AI supercomputers have doubled every 13 months

By Konstantin F. Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim

The US hosts the majority of GPU cluster performance, followed by China
Data Insight
Jun. 5, 2025
The US hosts the majority of GPU cluster performance, followed by China

By Konstantin F. Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim

Private-sector companies own a dominant share of GPU clusters
Data Insight
Jun. 5, 2025
Private-sector companies own a dominant share of GPU clusters

By Konstantin F. Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim

Power requirements of leading AI supercomputers have doubled every 13 months
Data Insight
Jun. 5, 2025
Power requirements of leading AI supercomputers have doubled every 13 months

By Konstantin F. Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim

Widespread adoption of new numeric formats took 3-4 years in past cycles
Data Insight
May 28, 2025
Widespread adoption of new numeric formats took 3-4 years in past cycles

By Venkat Somala and Luke Emberson

The computational performance of leading AI supercomputers has doubled every nine months
Data Insight
Updated Jun. 5, 2025
The computational performance of leading AI supercomputers has doubled every nine months

By Konstantin F. Pilz, Robi Rahman, James Sanders, Luke Emberson, and Lennart Heim

LLM responses to benchmark questions are getting longer over time
Data Insight
Apr. 17, 2025
LLM responses to benchmark questions are getting longer over time

By Luke Emberson, Ben Cottier, Josh You, Tom Adamczewski, and Jean-Stanislas Denain

The combined revenues of leading AI companies grew by over 9x in 2023-2024
Data Insight
Apr. 3, 2025
The combined revenues of leading AI companies grew by over 9x in 2023-2024

By Ben Snodin, David Owen, and Luke Emberson

LLM inference prices have fallen rapidly but unequally across tasks
Data Insight
Mar. 12, 2025
LLM inference prices have fallen rapidly but unequally across tasks

By Ben Cottier, Ben Snodin, David Owen, and Tom Adamczewski

Leading AI chip designs are used for around four years in frontier training
Data Insight
Mar. 5, 2025
Leading AI chip designs are used for around four years in frontier training

By Luke Emberson, Ben Snodin, and David Owen

Biology AI models are scaling 2-4x per year after rapid growth from 2019-2021
Data Insight
Feb. 21, 2025
Biology AI models are scaling 2-4x per year after rapid growth from 2019-2021

By Pablo Villalobos and David Atanasov

The stock of computing power from NVIDIA chips is doubling every 10 months
Data Insight
Feb. 13, 2025
The stock of computing power from NVIDIA chips is doubling every 10 months

By Luke Emberson and David Owen

Over 30 AI models have been trained at the scale of GPT-4
Data Insight
Updated Jun. 6, 2025
Over 30 AI models have been trained at the scale of GPT-4

By Robi Rahman, Lovis Heindrich, David Owen, and Luke Emberson

Chinese language models have scaled up more slowly than their global counterparts
Data Insight
Jan. 22, 2025
Chinese language models have scaled up more slowly than their global counterparts

By Ben Cottier

Frontier open models may surpass 1e26 FLOP of training compute before 2026
Data Insight
Jan. 15, 2025
Frontier open models may surpass 1e26 FLOP of training compute before 2026

By Luke Emberson

Training compute growth is driven by larger clusters, longer training, and better hardware
Data Insight
Jan. 8, 2025
Training compute growth is driven by larger clusters, longer training, and better hardware

By Luke Emberson and David Owen

Accuracy increases with estimated training compute
Data Insight
Updated Feb. 7, 2025
Accuracy increases with estimated training compute

By Jean-Stanislas Denain

Models with downloadable weights currently lag behind the top-performing models
Data Insight
Updated Feb. 7, 2025
Models with downloadable weights currently lag behind the top-performing models

By Jean-Stanislas Denain

US models currently outperform non-US models
Data Insight
Updated Feb. 7, 2025
US models currently outperform non-US models

By Jean-Stanislas Denain

Performance improves 13x when switching from FP32 to tensor-INT8
Data Insight
Oct. 23, 2024
Performance improves 13x when switching from FP32 to tensor-INT8

By Robi Rahman and David Owen

Leading ML hardware becomes 40% more energy-efficient each year
Data Insight
Oct. 23, 2024
Leading ML hardware becomes 40% more energy-efficient each year

By Robi Rahman

The NVIDIA A100 has been the most popular hardware for training notable machine learning models
Data Insight
Oct. 23, 2024
The NVIDIA A100 has been the most popular hardware for training notable machine learning models

By Robi Rahman

The computational performance of machine learning hardware has doubled every 2.3 years
Data Insight
Oct. 23, 2024
The computational performance of machine learning hardware has doubled every 2.3 years

By Robi Rahman

Performance per dollar improves around 30% each year
Data Insight
Oct. 23, 2024
Performance per dollar improves around 30% each year

By Robi Rahman

AI training cluster sizes increased by more than 20x since 2016
Data Insight
Oct. 23, 2024
AI training cluster sizes increased by more than 20x since 2016

By Robi Rahman

Leading AI companies have hundreds of thousands of cutting-edge AI chips
Data Insight
Oct. 9, 2024
Leading AI companies have hundreds of thousands of cutting-edge AI chips

By Josh You and David Owen

The power required to train frontier AI models is doubling annually
Data Insight
Sep. 19, 2024
The power required to train frontier AI models is doubling annually

By Luke Emberson and Robi Rahman

The length of time spent training notable models is growing
Data Insight
Aug. 16, 2024
The length of time spent training notable models is growing

By Luke Emberson

Training compute has scaled up faster for language than vision
Data Insight
Jun. 19, 2024
Training compute has scaled up faster for language than vision

By Robi Rahman and David Owen

The training compute of notable AI models has been doubling roughly every six months
Data Insight
Jun. 19, 2024
The training compute of notable AI models has been doubling roughly every six months

By Robi Rahman and David Owen

Training compute costs are doubling every eight months for the largest AI models
Data Insight
Jun. 19, 2024
Training compute costs are doubling every eight months for the largest AI models

By Ben Cottier and Robi Rahman

The size of datasets used to train language models doubles approximately every six months
Data Insight
Jun. 19, 2024
The size of datasets used to train language models doubles approximately every six months

By Robi Rahman and David Owen

Almost half of large-scale models have published, downloadable weights
Data Insight
Jun. 19, 2024
Almost half of large-scale models have published, downloadable weights

By Ben Cottier, Josh You, and Natalia Martemianova

The pace of large-scale model releases is accelerating
Data Insight
Jun. 19, 2024
The pace of large-scale model releases is accelerating

By Robi Rahman

Most large-scale models are developed by US companies
Data Insight
Jun. 19, 2024
Most large-scale models are developed by US companies

By Robi Rahman

Language models compose the large majority of large-scale AI models
Data Insight
Jun. 19, 2024
Language models compose the large majority of large-scale AI models

By Robi Rahman and Josh You