Epoch AI

📋Page Status

Page Type:ContentStyle Guide →Standard knowledge base article

Quality:51 (Adequate)⚠️

Importance:42 (Reference)

Last edited:2026-01-29 (3 days ago)

Words:4.6k

Backlinks:2

Structure:

📊 30📈 1🔗 6📚 66•11%Score: 14/15

LLM Summary:Epoch AI maintains comprehensive databases tracking 3,200+ ML models showing 4.4x annual compute growth and projects data exhaustion 2026-2032. Their empirical work directly informed EU AI Act's 10^25 FLOP threshold and US EO 14110, with their Epoch Capabilities Index showing ~90% acceleration in AI progress since April 2024.

Issues (2):

• QualityRated 51 but structure suggests 93 (underrated by 42 points)
• Links15 links could use <R> components

See also:EA Forum

Quick Assessment

Dimension	Assessment	Evidence
Research Impact	Very High	Cited in US AI Executive Order 14110, EU AI Act 10^25 FLOP threshold, Congressional testimony
Data Quality	Exceptional	3,200+ ML models tracked from 1950-present; most comprehensive public dataset
Methodology Rigor	High	Peer-reviewed publications (arXiv:2202.05924); transparent compute estimation methodology
Policy Influence	Strong	UK DSIT collaboration; JRC European Commission consultations; House of Lords evidence submission
Industry Usage	Widespread	OpenAI commissioned FrontierMath; Google DeepMind collaborated on ECI methodology
Funding	Stable	≈$7M through 2025 via Coefficient GivingCoefficient GivingCoefficient Giving (formerly Open Philanthropy) has directed $4B+ in grants since 2014, including $336M to AI safety (~60% of external funding). The organization spent ~$50M on AI safety in 2024, w...Quality: 55/100 grants
Team Size	≈34 employees	Founded by 7 researchers; now includes ML, economics, statistics, policy backgrounds
Key Metrics	Quantified	15M+ H100-equivalents tracked; 30+ models above 10^25 FLOP; 7-month doubling time for AI compute

Organization Details

Attribute	Details
Full Name	Epoch AI
Founded	April 2022
Location	San Francisco, CA (headquarters); remote-first operations
Status	Independent 501(c)(3) nonprofit (since early 2025; previously fiscally sponsored by Rethink Priorities)
Website	epoch.ai
Director	Jaime Sevilla (Mathematics and Computer Science background)
Key Outputs	ML Trends Database, Epoch Capabilities Index, FrontierMath Benchmark, GATE Economic Model, AI Chip Sales Tracker
Primary Funders	Coefficient GivingCoefficient GivingCoefficient Giving (formerly Open Philanthropy) has directed $4B+ in grants since 2014, including $336M to AI safety (~60% of external funding). The organization spent ~$50M on AI safety in 2024, w...Quality: 55/100 ($6.3M+ in grants), Carl Shulman ($100K), individual donors
GitHub	epoch-research

Overview

Epoch AI is a research institute dedicated to tracking and forecasting AI development through rigorous empirical analysis. Founded in April 2022 by Jaime Sevilla and six co-founders, Epoch has become the authoritative source for data on AI training compute, model parameters, hardware capabilities, and development timelines. Their research directly informs policy discussions, corporate planning, and academic research on AI trajectories. The New York Times praised their work for bringing “much-needed rigor and empiricism to an industry that often runs on hype and vibes,” and featured them in their 2024 Good Tech Awards.

The organization’s core contribution is maintaining comprehensive databases that enable quantitative analysis of AI progress. Their public database tracks over 3,200 machine learning models from 1950 to present, documenting the training compute, parameters, and capabilities of each system. By cataloging these metrics, Epoch provides the empirical foundation for discussions about AI timelines, resource constraints, and capability trajectories. Their work bridges the gap between speculative AI forecasting and evidence-based analysis.

Epoch’s research has been directly cited in major policy documents including the EU AI Act (which adopted their 10^25 FLOP compute threshold) and US Executive Order 14110. Their data informs Congressional hearings, and leading AI labs use their metrics for planning. As director Jaime Sevilla stated, “We want to do something similar for artificial intelligence to what William Nordhaus, the Nobel laureate, did for climate change. He set the basis for rigorous study and thoughtful action guided by evidence.” The organization represents a critical piece of epistemic infrastructure for understanding where AI development is headed and what constraints may shape its trajectory.

History and Evolution

Founding (2022)

Epoch AI emerged from a collaborative research effort that began when Jaime Sevilla, a Spanish researcher, put his Ph.D. on pause and issued a call for volunteers to systematically document the critical inputs of every significant AI model ever created. The initial team that responded went on to become Epoch’s founding members: Jaime Sevilla, Tamay Besiroglu, Lennart Heim, Pablo Villalobos, Eduardo Infante-Roldan, Marius Hobbhahn, and Anson Ho. Collectively, they brought backgrounds in Machine Learning, Statistics, Economics, Forecasting, Physics, Computer Engineering, and Software Engineering.

During Epoch’s first retreat in April 2022, the members decided to formalize as an organization and chose the name “Epoch” through a Twitter poll. When they published their findings in the paper “Compute Trends Across Three Eras of Machine Learning” in early 2022, the reaction was overwhelmingly positive, with the paper going viral in AI research communities. The paper documented the 10-billion-fold increase in training compute since 2010 and identified three distinct eras of ML development with different scaling dynamics.

Growth and Fiscal Sponsorship (2022-2024)

From its founding, Epoch was fiscally sponsored and operationally supported by Rethink Priorities, whose Special Projects team provided critical infrastructure for the growing organization. At founding, Epoch had a staff of 13 people (9 FTEs). Coefficient GivingCoefficient GivingCoefficient Giving (formerly Open Philanthropy) has directed $4B+ in grants since 2014, including $336M to AI safety (~60% of external funding). The organization spent ~$50M on AI safety in 2024, w...Quality: 55/100 (then Open Philanthropy) provided early support with a $1.96M grant for general support, followed by additional grants totaling over $6M through 2025.

During this period, Epoch expanded its database from the initial 123 models documented in their founding paper to over 3,200 models. They launched new data products including the AI Chip Sales tracker, Parameter Counts database, and AI Supercomputer Tracker. The team grew to approximately 34 employees with headquarters established in San Francisco.

Independence and Expansion (2025-Present)

In early 2025, Epoch spun out from its fiscal sponsor and began operating as an independent 501(c)(3) nonprofit organization. This transition marked their maturation from a research project to a fully independent institution. Key 2025 developments included:

• Launch of the Epoch Capabilities Index (ECI) in October 2025, a unified metric combining scores from 37 benchmarks
• Completion of FrontierMath Tier 4, commissioned by OpenAI, featuring 50 research-level mathematics problems
• Publication of more plots and visualizations in 2025 than in all previous years combined
• Development of the GATE model for forecasting AI’s economic impact

Team

Leadership and Key Researchers

Name	Role	Background	Key Contributions
Jaime Sevilla	Director	Mathematics, Computer Science	Founded Epoch; leads research on AI forecasting and trends
Tamay Besiroglu	Research Advisor (former co-founder)	Economics of computing	Co-authored founding paper; now leads Mechanize startup
Lennart Heim	Co-founder	Computer Engineering	Compute governance research; hardware tracking
Pablo Villalobos	Researcher	Statistics	Data constraints research; “Will We Run Out of Data?” paper
Marius Hobbhahn	Co-founder	ML, Physics	Compute trends analysis
Anson Ho	Co-founder	ML, Software Engineering	Database development; trend analysis
Eduardo Infante-Roldan	Co-founder	Economics	Economic modeling

Team Composition

The current team of approximately 34 employees includes researchers with diverse backgrounds:

• Machine Learning researchers: Core technical expertise in model architectures and training
• Economists: Analyze AI’s economic implications and build forecasting models
• Statisticians: Develop rigorous methodologies for trend analysis
• Policy analysts: Translate research findings for governance contexts
• Data engineers: Maintain and expand the organization’s databases

Key Research Areas

Loading diagram...

Training Compute Estimation Methodology

A critical contribution of Epoch AI is their rigorous methodology for estimating the compute used to train machine learning models. This methodology enables accurate comparisons across models and time periods, forming the foundation for their scaling analysis.

Two Primary Estimation Strategies

Epoch uses two complementary approaches to estimate training compute:

Method	Inputs Required	When Used	Precision
Architecture-based	Model architecture, training data size, parameter count	When architecture details are published	High
Hardware-based	GPU type, training time, utilization rate	When hardware details are available	Moderate

Architecture-based estimation: Epoch maintains detailed tables of common neural network layers, estimating parameters and FLOP per forward pass. For many layers, forward pass FLOP approximately equals twice the parameter count. A backward pass adds approximately 2x the forward pass FLOP, yielding the common heuristic: total training FLOP approximately equals 6 times parameters times tokens.

Hardware-based estimation: When architecture details are unavailable, Epoch calculates compute from GPU training time multiplied by peak GPU performance, adjusted by utilization rate. Their empirical analysis found utilization rates typically range from 0.3 to 0.75 depending on architecture and batch size.

Key Methodological Insights

• The backward/forward FLOP ratio is “very likely 2:1” after correcting for common counting errors
• The “Theory method” multiplies forward pass FLOP by 3.0 to account for backward pass
• Larger batch sizes yield more consistent utilization rates
• Parameter sharing (as in CNNs) and word embeddings require special handling

Training Compute Trends

Metric	Finding	Time Period	Source
Compute Growth	4.4x per year	2010-2025	Epoch Trends
Doubling Time	≈5-6 months	Deep Learning era (2012+)	arXiv:2202.05924
Pre-Deep Learning Doubling	≈20 months	Before 2010	Moore’s Law trajectory
Models above 10^25 FLOP	30+ models from 12 developers	As of June 2025	Epoch Data
Global AI Chip Capacity	15M+ H100-equivalents	2025	AI Chip Sales

Three Eras of Machine Learning

Epoch’s foundational 2022 paper identified three distinct eras with different compute scaling dynamics:

Era	Period	Doubling Time	Characteristics
Pre-Deep Learning	Before 2010	≈20 months	Followed Moore’s Law; academic-dominated
Deep Learning	2010-2015	≈5-6 months	Rapid scaling; breakthrough architectures
Large-Scale	2015-present	≈5-6 months	2-3 OOM more compute than previous trend; industry-dominated

This analysis corrected earlier estimates (Amodei and Hernandez 2018) that suggested 3.4-month doubling, finding the actual rate closer to 5-6 months with approximately 10x more data points.

Historical Compute Scaling

Epoch’s database reveals the dramatic scaling of AI training compute:

Era	Representative Model	Training Compute (FLOP)	Approximate Cost
2012	AlexNet	10^17	Thousands
2017	Transformer (original)	10^18	Tens of thousands
2020	GPT-3	10^23	Millions
2023	GPT-4	10^25	Tens of millions
2024	Frontier models	10^26	≈$100 million
2027 (proj.)	Next-gen frontier	10^27+	Greater than $1 billion

The first model trained at the 10^25 FLOP scale was GPT-4, released in March 2023. As of June 2025, Epoch has identified over 30 publicly announced AI models from 12 different developers that exceed this threshold.

Hardware Trends

Metric	Annual Growth	Current Status
GPU FLOP/s (FP32)	1.35x	Continuing Moore’s Law trajectory
GPU FLOP/s (FP16)	Similar to FP32	Optimized for ML workloads
NVIDIA Total Compute	2.3x since 2019	Hopper generation: 77% of total
Global AI Compute Capacity	3.3x per year (7-month doubling)	15M+ H100-equivalents total

AI Chip Sales Database

Epoch’s AI Chip Sales data explorer is the most comprehensive public dataset tracking global AI compute capacity across vendors:

Vendor	Coverage	Key Metrics Tracked
NVIDIA	Primary	GPU sales, FLOP capacity, power consumption
Google (TPU)	Included	Custom silicon production
Amazon (Trainium)	Included	Cloud AI accelerators
AMD	Included	MI series GPUs
Huawei	Included	Ascend chips (domestic China)

Key finding: Global computing capacity has been growing by 3.3x per year, equivalent to a doubling time of approximately 7 months.

AI Supercomputer Projections

Epoch’s analysis of AI supercomputer trends projects significant scaling challenges:

Year	Projected Chips	Estimated Cost	Power Required
2024	≈100,000	≈$10 billion	≈1 GW
2027	≈500,000	≈$50 billion	≈4 GW
2030	≈2 million	≈$200 billion	≈9 GW

The 9 GW power requirement for 2030 frontier training represents the equivalent of 9 nuclear reactors—a scale beyond any existing industrial facility. This represents a potential binding constraint on AI scaling.

Geographic Distribution of AI Compute

Region	Share of AI Supercomputer Capacity	Trend
United States	≈75%	Dominant and growing
China	≈15%	Second place, facing chip restrictions
Europe	≈5%	Limited domestic capacity
Other	≈5%	Emerging efforts

The shift from academic to industry dominance has been dramatic:

Year	Industry Share	Academic/Government Share
2019	≈40%	≈60%
2022	≈65%	≈35%
2025	≈80%	≈20%

Data Constraints Research

Epoch’s influential research on training data constraints (the “data wall”) has become central to discussions of AI scaling limits. Their paper “Will We Run Out of ML Data?” projects when AI development may exhaust human-generated training data.

Key Projections

Data Source	Current Status	Exhaustion Projection (80% CI)
Public web text	Heavily utilized	2026-2028
Books and academic papers	Largely incorporated	2027-2030
All human-generated text	Approaching limits	2026-2032

The exact date depends on scaling assumptions. According to researcher Tamay Besiroglu: “There is a serious bottleneck here. If you start hitting those constraints about how much data you have, then you can’t really scale up your models efficiently anymore. And scaling up models has been probably the most important way of expanding their capabilities.”

Overtraining Factor Analysis

Overtraining Factor	Data Exhaustion Year	Example
Compute-optimal (1x)	≈2028	Enough for 5x10^28 FLOP model
5x overtrained	≈2027	Common practice
10x overtrained	≈2026-2027	Llama 3-70B level
100x overtrained	≈2025	Extreme efficiency

Updated Estimates (2025)

Epoch’s analysis has evolved based on new evidence:

• The effectiveness of carefully filtered web data and multi-epoch training has substantially increased estimates of available high-quality data
• After accounting for data quality, availability, multiple epochs, and multimodal tokenizer efficiency, Epoch estimates 400 trillion to 20 quadrillion tokens available for training by 2030
• This allows for training runs from 6x10^28 to 2x10^32 FLOP

Mitigating Factors

Epoch identifies three categories of innovation that could extend the scaling runway:

Mitigation	Mechanism	Status
Synthetic data	AI-generated training data	Active research; quality concerns remain
Multimodal data	Images, video, audio expand data pool	Increasingly used
Data efficiency	Better algorithms require less data	Ongoing improvements

While Sam Altman noted OpenAI experiments with “generating lots of synthetic data,” he expressed reservations: “There’d be something very strange if the best way to train a model was to just generate, like, a quadrillion tokens of synthetic data and feed that back in.” Research shows training on AI-generated data can produce “model collapse” with degraded outputs.

Epoch Capabilities Index (ECI)

The Epoch Capabilities Index (ECI), launched in October 2025, represents a major methodological advance in measuring AI progress. As individual benchmarks saturate, ECI provides a unified scale for comparing models across time.

Methodology

ECI combines scores from 37 distinct benchmarks into a single “general capability” scale, similar to how IQ tests capture broad underlying capability:

Aspect	Details
Benchmarks included	37 distinct benchmarks
Evaluations used	1,123 distinct evaluations
Models covered	147 models
Time span	December 2021 - December 2025
Methodology basis	”A Rosetta Stone for AI Benchmarks” (collaboration with Google DeepMind AGI Safety team)

ECI scores function like Elo ratings: absolute values are less meaningful than relative comparisons. The scale is linear, so a 10-point jump should be equally significant whether moving from 100 to 110 or from 140 to 150.

Key Finding: 90% Acceleration

Epoch’s analysis reveals a significant acceleration in AI capabilities progress:

Period	Annual ECI Growth	Key Driver
December 2021 - April 2024	≈8 points/year	Scaling laws, architecture improvements
April 2024 - December 2025	≈15 points/year	Reasoning models, reinforcement learning
Acceleration	≈90%	Coincides with rise of o1-style reasoning models

This acceleration is corroborated by METR’s Time Horizon benchmark, which found a ~40% acceleration in task completion capabilities starting around the same period.

FrontierMath Benchmark

FrontierMath is Epoch’s benchmark of original, expert-crafted mathematics problems designed to evaluate advanced reasoning capabilities—problems that typically require hours or days for expert mathematicians to solve.

Development and Structure

Aspect	Details
Total problems	350 (300 base + 50 Tier 4 expansion)
Collaborating mathematicians	60+ from leading institutions
Notable contributors	14 International Mathematical Olympiad gold medalists, 1 Fields Medal recipient
Problem domains	Computational number theory, abstract algebraic geometry, and other advanced fields
Commissioning	OpenAI commissioned the core 300 problems

Tier Structure

Tier	Problem Count	Difficulty	Typical Human Solve Time
Tier 1	≈100	Advanced undergraduate	Hours
Tier 2	≈100	Graduate level	Hours to days
Tier 3	≈100	Research level	Days
Tier 4	50	Short research projects	Days to weeks

Model Performance

While leading AI models achieve near-perfect scores on traditional math benchmarks (GSM-8k, MATH), FrontierMath reveals substantial gaps:

Model	FrontierMath Score	Traditional Benchmarks	Notes
GPT-4o, Claude 3.5	Less than 2%	Greater than 90% on MATH	Baseline frontier models
o3 (December 2024)	≈25% (announced)	Near-perfect on MATH	Pre-release version
o3 (April 2025 release)	≈10%	Near-perfect on MATH	Official Epoch evaluation

The discrepancy between o3’s announced 25% and measured 10% reflects differences in model versions and benchmark composition over time. Both the model and benchmark changed between December 2024 and April 2025.

Significance

FrontierMath addresses two critical challenges:

1. Benchmark saturation: Traditional math benchmarks no longer differentiate frontier models
2. Data contamination: Using entirely new, unpublished problems with automated verification

GATE Economic Model

The GATE model (Growth and AI Transition Endogenous) is Epoch’s integrated assessment model of AI’s economic impact, published in 2025 (arXiv:2503.04941).

Core Dynamics

GATE models an automation feedback loop: investments drive increases in compute for training and deploying AI, which leads to gradual task automation, which generates returns enabling further investment.

Key Predictions

Metric	GATE Projection	Context
AI Investment Peak	Greater than 10% of world GDP	≈50x increase over current levels
Growth at 30% automation	Greater than 20% annual GWP growth	Comparable to industrial revolution peaks
Growth at 40% automation	≈12% annual GWP growth	Comparable to East Asian miracle economies

Interpretation Caveats

Epoch explicitly cautions against treating GATE outputs as precise quantitative predictions. The model illustrates key dynamics rather than providing forecasts. According to their analysis: “These findings suggest that those who are confidently either extremely skeptical or extremely bullish about an unprecedented growth acceleration due to AI are likely miscalibrated.”

A public GATE playground allows users to modify parameters and explore scenarios.

Policy Impact

Epoch’s research has directly influenced major AI governance frameworks. Their compute trend data provides the empirical foundation for regulatory thresholds.

Compute Thresholds in Regulation

Policy	Threshold	Epoch’s Role
EU AI Act	10^25 FLOP for systemic risk models	Epoch data cited in JRC technical documents
US Executive Order 14110	10^26 FLOP for reporting requirements	Threshold informed by Epoch trend analysis
UK Frontier AI Safety	Uses compute as capability proxy	Methodology collaboration with UK DSIT

The EU AI Act explicitly references the statistical relationship between training compute and model capabilities documented by Epoch, noting that “performance of 231 language models (measured in log-perplexity) against scale (measured in FLOP)” shows clear trends.

Government Engagements

Government Body	Engagement Type	Year
UK DSIT	Consultation on “Frontier AI: capabilities and risks”	2023
JRC European Commission	Collaboration on AI Act technical documentation	2023-2024
House of Lords	Evidence submission on language models	2023
NIST	Input on AI Risk Management Framework	2023-2024
US OSTP	Briefings on compute trends	2023-2024

Model Count Projections

Epoch’s analysis of how many models will exceed compute thresholds directly informs regulatory planning:

Threshold	Models (June 2025)	Developers
10^23 FLOP	Hundreds	Dozens
10^25 FLOP	30+	12
10^26 FLOP	Several	Major labs

Key Publications

Publication	Year	Key Finding	Citation
”Compute Trends Across Three Eras of Machine Learning”	2022	4.4x annual growth; 5-6 month doubling; three distinct eras	arXiv:2202.05924
”Will We Run Out of ML Data?“	2022	Data exhaustion projected 2026-2032	Epoch Blog
”Estimating Training Compute of Deep Learning Models”	2022	Methodology for FLOP estimation	Epoch Blog
”Can AI Scaling Continue Through 2030?“	2024	Analysis of compute, data, energy constraints	Epoch Blog
”AI Capabilities Progress Has Sped Up”	2024	≈90% acceleration since April 2024	Epoch Data Insights
”FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI”	2024	Frontier models solve less than 2%	arXiv:2411.04872
”GATE: An Integrated Assessment Model for AI Automation”	2025	Economic modeling of AI transition	arXiv:2503.04941
”How Well Did Forecasters Predict 2025 AI Progress?“	2025	Metacognitive evaluation of forecasting	Epoch Blog
”Global AI Computing Capacity is Doubling Every 7 Months”	2025	15M+ H100-equivalents; 3.3x annual growth	Epoch Data Insights

Funding

Epoch AI has raised approximately $7 million through September 2025, primarily from Coefficient GivingCoefficient GivingCoefficient Giving (formerly Open Philanthropy) has directed $4B+ in grants since 2014, including $336M to AI safety (~60% of external funding). The organization spent ~$50M on AI safety in 2024, w...Quality: 55/100 grants.

Coefficient Giving Grants

Grant	Amount	Purpose	Date
General Support (2022)	$1,960,000	Initial organizational support	2022
General Support (2023)	$4,132,488	Two-year general support	2023
Worldview Investigations	$188,558	AI-related worldview research	2023
General Support (2025)	Undisclosed	Independent operations	2025

Other Funding

• Carl Shulman: $100,000 individual donation
• Various individual donors
• Contract revenue from clients including AI labs and government offices

Coefficient Giving has cited Epoch as producing “world-class work that is widely read, used, and shared.”

Comparison with Other Forecasting Organizations

Organization	Focus	Methodology	Key Strength	Compute Expertise
Epoch AI	Empirical AI trends	Database analysis, benchmark development	Hardware/compute tracking; 3,200+ models	Primary focus
MetaculusOrganizationMetaculusMetaculus is a reputation-based forecasting platform with 1M+ predictions showing AGI probability at 25% by 2027 and 50% by 2031 (down from 50 years away in 2020). Analysis finds good short-term ca...Quality: 50/100	Crowd forecasting	Prediction aggregation	Diverse questions; large forecaster base	Questions only
Our World in Data	Data visualization	Curates authoritative sources	Broad topic coverage; accessibility	Uses Epoch data
AI Impacts	AI forecasting	Expert surveys, trend extrapolation	Timeline estimates	Moderate
QURIOrganizationQURI (Quantified Uncertainty Research Institute)QURI develops Squiggle (probabilistic programming language with native distribution types), SquiggleAI (Claude-powered model generation producing 100-500 line models), Metaforecast (aggregating 2,1...Quality: 48/100	Epistemic tools	Software development	Probabilistic modeling	Limited

Our World in Data directly incorporates Epoch’s compute trend data in their AI visualizations, extending Epoch’s reach to a broader audience.

Strengths and Limitations

Strengths

Strength	Evidence
Comprehensive data	Most complete public database: 3,200+ ML models from 1950-present
Transparent methodology	Open documentation of compute estimation methods; peer-reviewed publications
Policy relevance	Directly cited in EU AI Act, US EO 14110; collaborations with UK DSIT, JRC
Regular updates	Databases continuously maintained; published more plots in 2025 than all previous years
Methodological innovation	ECI provides unified capability measurement; FrontierMath addresses benchmark saturation
Industry recognition	New York Times “Good Tech Awards” 2024; praised for “rigor and empiricism”

Limitations

Limitation	Implication
Historical focus	Primarily backward-looking; projections carry significant uncertainty
Compute-centric	Algorithmic efficiency improvements harder to quantify than hardware scaling
Industry opacity	Labs don’t disclose training details; estimates rely on public information
Threshold arbitrariness	10^25 FLOP thresholds are useful proxies but don’t directly measure capability
US-centric	Limited visibility into Chinese AI development due to information barriers
Funding concentration	Heavy reliance on Coefficient Giving creates potential dependency

Critical Assessment

What Epoch Does Well

Epoch fills a crucial gap in the AI ecosystem by providing rigorous empirical grounding for discussions that previously relied on intuition and speculation. Before Epoch, claims about AI progress rates were often based on anecdotes or marketing materials. Epoch’s systematic data collection enables evidence-based analysis of trends that matter for policy and planning.

Their methodology for compute estimation has become the de facto standard, cited by academics, policymakers, and industry alike. The FrontierMath benchmark addresses a genuine problem—traditional benchmarks saturating—with a thoughtful approach using novel problems from credentialed experts.

Key Uncertainties

Key Questions (6)

• How reliable are compute estimates for models where labs don't disclose training details?
• Will the historical relationship between compute and capabilities continue, or are we approaching diminishing returns?
• Can policy thresholds based on compute (10^25 FLOP) remain meaningful as algorithmic efficiency improves?
• How should Epoch's projections be weighted against insider knowledge from labs?
• Will data constraints prove as binding as projected, or will synthetic data and efficiency gains extend the runway?
• Can a small organization maintain comprehensive coverage as AI development accelerates globally?

Perspectives on Epoch’s Role

Views on Epoch AI's Contribution

Value and Limitations of Empirical AI Tracking (3 perspectives)

Essential Infrastructure

Dominant

Epoch provides irreplaceable empirical grounding for AI policy and research. Their data and analysis have elevated discourse from speculation to evidence-based discussion. Expansion of their work should be a priority.

AI governance researchers · Policymakers · Safety-focused AI researchers

Valuable but Incomplete

Moderate

Epoch's compute tracking is useful but overemphasizes hardware relative to algorithms and data quality. Capability improvements from techniques like RLHF and chain-of-thought are harder to quantify but equally important.

ML researchers · Industry practitioners

Projection Skepticism

Minority

Historical trends are useful for backward-looking analysis but extrapolation to future capabilities is unreliable. Past growth rates may not predict discontinuities or saturation.

Forecasting skeptics · Some AI researchers

Timeline

Date	Event
April 2022	Epoch AI founded; team of 7 researchers; fiscally sponsored by Rethink Priorities
February 2022	”Compute Trends Across Three Eras” published; paper goes viral in AI research community
2022	First Coefficient Giving (then Open Philanthropy) grant ($1.96M)
2022	”Will We Run Out of ML Data?” published; establishes data wall projections
2023	Database grows to 800+ models; UK DSIT and JRC collaborations begin
2023	Additional Coefficient Giving (then Open Philanthropy) grants ($4.3M); House of Lords evidence submission
2024	FrontierMath benchmark developed with 60+ mathematicians; OpenAI commissions 300 problems
2024	Database exceeds 3,200 models; AI Chip Sales tracker launched
2024	New York Times “Good Tech Awards” recognition
October 2024	Epoch Capabilities Index (ECI) launched with 37 benchmarks
December 2024	o3 announced with 25% FrontierMath score (later measured at ≈10%)
Early 2025	Spin out from fiscal sponsor to independent 501(c)(3)
2025	GATE economic model published; 15M+ H100-equivalents tracked
2025	Published more visualizations than all previous years combined

Sources and External Links

Official Resources

• Epoch AI Website
• ML Trends Dashboard
• Epoch AI Database
• Epoch Substack
• 2025 Impact Report
• GitHub: epoch-research
• Team Page
• Funding Information

Key Data Products

• AI Models Database - 3,200+ models from 1950-present
• AI Chip Sales - Global compute capacity tracking
• Epoch Capabilities Index - Unified capability measurement
• FrontierMath Benchmark - Advanced mathematical reasoning
• GATE Model Playground - Interactive economic modeling

Academic Publications

• Compute Trends Across Three Eras (arXiv:2202.05924)
• FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI (arXiv:2411.04872)
• GATE: An Integrated Assessment Model for AI Automation (arXiv:2503.04941)

Methodology Documentation

• Estimating Training Compute
• How to Measure FLOP Empirically
• Interpreting the Epoch Capabilities Index

Coefficient Giving Grants

• Epoch General Support (2022)
• Epoch AI General Support (2025)
• Epoch AI Worldview Investigations

Policy Documents Citing Epoch

• EU AI Act Technical Documentation (JRC)
• Training Compute Thresholds in AI Regulation (arXiv:2405.10799)

Media Coverage

• The Epoch AI Researcher Trying to Glimpse the Future of AI (TIME)
• Announcing Epoch (EA Forum)