Epoch AI 2025 impact report

In 2025, we saw AI continue to increase in scale and importance. AI companies reached annual revenues totalling tens of billions of dollars, and are building data centers that individually cost comparable amounts. Leading benchmarks show capabilities accelerating, propped up by the establishment of reasoning models, such as OpenAI’s oN model series. And we have seen an incredible diffusion of capabilities, with Chinese open weight models such as DeepSeek R1 closing in the gap with US frontier models released only months before.

Epoch AI has responded with new and expanded initiatives to advance its mission of sharing up-to-date information about – and making sense of – the trajectory of AI. We are excited to share a recap of our work in 2025, and our plans for 2026.

We are raising $3 million to execute a more ambitious version of our plans. Donations can be made directly through our website. For those considering a substantial contribution, or commissioning a project, please contact us at donate@epoch.ai.

Highlights from 2025

AI infrastructure became a major focus of investment and public attention in 2025. We pursued two related initiatives, starting with the creation of the GPU Clusters Data Explorer (originally called AI Supercomputers), followed by the ongoing build-out of the Frontier Data Centers Data Explorer, using satellite and permit data to track compute, power use, and construction timelines.

Why this matters: AI compute has long been an important input to AI capabilities. It is essential not only to enable large scale training, but also to conduct the experiments that lead to further progress. Tracking the construction of large data centers provides early, concrete signals about how quickly AI development capacity is expanding—and where it may run into limits.

Early in 2025, we launched a revamped version of our Benchmarking hub. Its landing page was our most visited page in 2025. Focused on top AI models, this page gathers evaluations reported by developers and third parties, as well as those run by Epoch.

As individual benchmarks saturate, it has become harder to compare frontier models using any single score. To address this, we introduced the Epoch Capabilities Index (ECI) in October, a composite metric that aggregates performance across multiple benchmarks to provide a more stable measure of model capability. The ECI combines at least four benchmark scores per model, drawing from over three dozen benchmarks in total, and performs well as a predictor of benchmark performance. This approach was developed as part of the “Rosetta Stone” collaboration with researchers from Google DeepMind.

Why this matters: Benchmark evaluations are one of the most straightforward – yet also ephemeral – ways to measure improvements in AI capabilities. Our Capabilities Index highlights the broader trend of improvements. For example, the ECI helped us identify a potential acceleration in AI capabilities near April 2024.

We completed and delivered FrontierMath Tier 4, a new tier of difficulty for our math benchmark, commissioned by OpenAI. Tier 4 consists of a collection of 50 research-level problems, including 2 public problems and a 20 question private holdout set. The problems were crafted by a team including world-leading university mathematics professors and postgraduate researchers, designed to test deep mathematical reasoning.

Most Tier 4 problems were designed or improved in a symposium attended by leading mathematicians, where problems were tested and approved by a panel of experts. Compared to FrontierMath Tiers 1-3, this has resulted in problems that are more difficult and harder to game with shortcuts, improving our ability to recognize genuinely strong mathematical reasoning that would impress a professional mathematician.

Why this matters: FrontierMath Tier 4’s focus on research-level problems allows us to track the capacity of new AI models to contribute to mathematical research. It is a benchmark that has remained largely unsaturated, with only 17 out of the 48 private questions solved across all models as of January 2026.

We released GATE, a framework for exploring how AI automation can affect the entire economy. GATE is a macroeconomic model that describes how investment in AI hardware and R\&D could lead to increased automation and productivity which enables further investments in automation. Our model illustrates how we could see explosive growth from AI, with over a fifth of the economy’s yearly output reinvested into AI, even under conditions of uncertainty about the total degree of automatability.

Why this matters: GATE is the most complete macroeconomic model we are aware of for the effects of AI automation. Further work would involve calibrating the model with real-world data, and fine-tuning the equations and parameters to produce realistic predictions of the AI trajectory. As a framework, it is already enabling economists and researchers to understand key dynamics of AI development.

During 2025, we responded to the increased tempo of AI developments by publishing shorter material on a weekly cadence through two distinct formats.

Our Data Insights are short, authoritative data investigations centered around a key graph and takeaway, meant to be accessible and citable sources for important AI trends. Popular Data Insights in 2025 include our analysis of inference price efficiency, of AI accessible on consumer hardware, and OpenAI’s allocation of compute between inference and development.

Our Gradient Updates newsletter offers leading-edge commentary and (when appropriate) speculative forays into important AI topics by individual authors, including some guest posts. Topics that we covered include a breakdown of the innovations introduced by DeepSeek v3, the energy costs of ChatGPT, and an analysis of how far reasoning models could scale.

Why this matters: Our audience is busy, and the AI industry moves fast. To help readers interpret our data efficiently, we publish concise analyses that highlight key insights and context, with timely commentary. These shorter formats complement our databases and longer reports.

In a report commissioned by Google DeepMind, we extrapolated existing trends in scaling compute, power, and data for training to understand the required inputs to maintain the current trend of progress to 2030. We also examined potential bottlenecks to scaling, and in each case found that they are likely to be surmountable.

We then extrapolated how this would affect performance in four domains: software engineering, mathematics, molecular biology, and weather prediction.

Why this matters: This report presents a core feature of how we currently understand the trajectory of AI: exponentially larger investments and inputs to development can lead to large advances in performance. This analysis then supports concrete extrapolations and insight into how such improved AI capabilities can affect science.

Epoch AI by the numbers

Outputs

Reach

Finances and organization

Press and citations

Our work has been extensively referenced by decision-makers and covered by the media. Below is a short selection of notable mentions, illustrating our research’s influence in the AI discourse.

Paid engagements

We entered several paid engagements with organizations from government, the AI industry, and investors to help them understand the future of AI and to complement our funding from donations.

On the benchmarking front, we released the aforementioned FrontierMath Tier 4, commissioned by OpenAI. We also started a collaboration funded by METR to develop a long-horizon software engineering benchmark. On model evaluations, xAI and Google DeepMind commissioned in-depth evaluations of the math capabilities of Grok 4 and Gemini Deep Think.

Google Deepmind also commissioned the AI in 2030 research report, as well as our co-authored “Rosetta Stone” paper. In the latter, we developed a method to aggregate and “translate” across benchmark scores for models evaluated on different subsets of benchmarks over multiple years. This collaboration helped us independently launch our Epoch Capabilities Index (ECI).

Other paid engagements include a report on the power demand for AI training, commissioned by the energy nonprofit EPRI, data insights commissioned by the AI Index and the Advanced Research and Invention Agency (ARIA), and consultations with UK AI Security Institute, EU AI Office, Sequoia Capital, and Bridgewater Associates. You can learn more about our consultancy services here.

Events and other engagements

Through the year we participated in a number of events and unpaid engagements to disseminate our research and inform the public about AI.

We gave briefings to U.S. congresspeople at the Aspen Institute Congressional AI Conference, to the UK AI directorate, and to Capitol Hill staffers, and we delivered talks for institutions such as Schmidt Sciences and the Institute for Progress.

We also participated in events including the EPRI annual conference on energy utilities, the Robotics: Science and Systems (RSS) conference, UK DSIT’s first AI Energy Council, the Global Hive Datacenter Summit, The Curve, and OpenAI’s Economic Research Conference.

Additionally, our staff were guests on prominent podcasts, including a16z and Dwarkesh, and we regularly engage with and are interviewed by journalists. Reach out to media@epoch.ai to arrange an interview, briefing, or talk.

Testimonials from our audience

We’re in regular dialogue with key stakeholders who benefit from our work. In our 2025 impact survey, we asked them to share about how our research has supported their work in the past year. Here are some of their responses.

"A key crux for CG's AI grantmaking strategy is AI timelines. As we've worked on an updated timelines analysis to inform that crux, Epoch reports have been a frequent source of parameter value estimates."

"Epoch is highly trusted by all camps. At least in DC federal policy circles, I’ve heard people say ‘I like what Epoch wrote’… You’re not seen as having ‘motivated reasoning’ in the same way that the heavily-silo-ed AI discourse is."

"We used the Epoch Capabilities Index to determine what the recent rate of software progress has been, which is input as a parameter into our upcoming timelines+takeoff model."

"Epoch is among the orgs whose work I cite most often."

"Epoch AI is probably the single organization I cite the most in my writing."

"Epoch's analysis and commentary—around, for example, the rise of reasoning models and the growing importance of test-time compute—has helped me thoughtfully and accurately characterize key trends in my policy analysis and reports."

"At Google DeepMind, we're using the Rosetta Stone approach to understand capabilities of our internal models, which supports a wide variety of decision-making. We're also doing some work to extend the approach to forecast dangerous capabilities; while I'm not yet sure whether we'll use this, it seems promising."

"I gave GPT 5 pro csv files of a bunch of your stats and had it turn them into a giant Anki deck. This has been very handy for refining my mental models and always having the right numbers on hand for doing fermi estimates and contextualizing news."

"We wrote a report on AI compute strategy in Germany and relied heavily on the GPU cluster database to compare compute capacity across Germany and the rest of the world. The report informed conversations with senior policymakers in Germany."

"I use Epoch materials in my talks regularly and my colleagues at the Stanford AI Index use it extensively. I am most excited about the insights being produced these days."

Governance and Transparency

In early 2025, Epoch AI spun out from its fiscal sponsor and began operating as an independent 501(c)(3) non-profit organization. This means that we now have our own board of directors, currently consisting of Tom Davidson, Ajeya Cotra, Jaime Sevilla, and Maria de la Lama.

In June, our director Jaime Sevilla published “What is Epoch AI?”, which lays out our mission to improve society’s understanding of the trajectory of AI and explains how that mission guides our choice of collaborators and projects. We also now publish detailed information about our funders, clients, collaborations, and conflicts of interest on our Funding page. These disclosures are part of our commitment to making our work legible and staying accountable to our audience.

Our plans for 2026

We expect the AI field will continue to grow in scale, revenue, and investment. And we plan to grow alongside it, so we can provide data and rigorous analysis to help society understand what is happening.

Our three key emphases for 2026 will be expanding our data programs; benchmark development & evaluations; and research & consultations. To make our research more accessible, we are also planning a series of initiatives to expand the reach of our website and other communications.

Curating and analyzing data is the backbone of our work at Epoch AI. We will continue expanding our data programs in 2026.

Expect an expansion of our Data Explorers for frontier data centers, AI companies, and benchmarking. We’ll also expand our work tracking compute across relevant actors, building on our first release of 2026: AI chip sales.

We’ll be building on the Epoch Capabilities Index: releasing specialized variants, benchmark difficulty ratings, and research into the impact of spending on capabilities.

Meanwhile, we will continue to maintain our ever-growing databases of AI models and AI hardware, revamping both to keep up with the rapid progress in the field. Our strategic priority will be the very frontier of AI: keeping up-to-date information especially on developments from top AI organizations globally, with less intense focus on non-frontier AI development.

Other data collection projects we are considering include: an expanded polling program to learn about AI usage, building on our prior work with Blue Rose; and data from earlier, upstream parts of the AI supply chain - such as fabs and semiconductor tools. We aim to quickly react to new sources of data, prioritizing the most important, under-served, and tractable sources of potential insight into the trajectory of AI.

The main ways we plan to share our data are through weekly Data Insights, Data Explorers, and the upcoming topic pages. Topic pages are one of our new flagship efforts to make our work easier to discover and understand - each topic page will collect our most important data, charts, publications, and other resources on a given topic into convenient and easily accessible overviews.

Understanding and monitoring trends in AI capabilities is critical to our broader mission. During 2025, we invested significantly in related projects: a major revamp of our benchmarking hub, the release of FrontierMath, and the development of the Epoch Capabilities Index as an aggregate metric of benchmark performance.

In 2026 we plan to continue expanding this work.

We will continue to build benchmarks filling gaps for economically and scientifically important AI capabilities. In 2026, we are already developing two new benchmarks. The first focuses on open mathematical problems whose solutions can be automatically verified. This allows us to track progress in mathematical reasoning even as FrontierMath Tier 4 saturates – and to test whether AI systems can contribute to advances at the research frontier. The second benchmark, developed in collaboration with METR, targets long-horizon software development tasks. These tasks go beyond bug fixes or optimization, assessing AI systems’ ability to implement complete software projects.

We also plan to continue evaluating models, both on benchmarks that others have created and our own. And because this is challenging to do for the wide range of models that we hope to measure, our plans include a significant investment into our underlying benchmarking infrastructure. Our goal is timely evaluation of all models relevant to tracking trends in capabilities.

Finally, we plan to continue producing qualitative reports about specific models and specific benchmarks. This includes deep-dives into individual model capabilities, going beyond benchmark scores to understand how models accomplish tasks and what failure modes remain. This also includes producing reviews of benchmarks themselves, in order to better understand how to interpret headline results.

Our work uniquely positions us to help key stakeholders understand critical developments in AI. We will continue to engage partners — including governments, think tanks, research nonprofits, investors, and model developers — through consultations and commissioned research.

Some topics we plan to investigate include: compute efficiency improvements; trends in robotics; indicators of AI enterprise adoption; the cost and composition of training data; and evidence around AI acceleration scenarios, such as the automation of AI R\&D or other parts of the AI supply chain.

We will also continue to offer big-picture syntheses of our work. For example, we plan to release an interactive model extrapolating key trends in revenue and compute through the next few years, exploring what we see as the “default” trajectory of AI going forward.

We expect to support discussion via commentary in our weekly Gradient Updates newsletter and by releasing podcast episodes.

We plan to continue improving how we present data and communicate our findings.

As part of our roadmap, we are planning a large visual update to our website. This will include topic pages that introduce many important trends in AI to a wide audience.

We released our first 5 podcasts in 2025, available in popular podcast platforms (including Apple Podcasts, Spotify, Amazon Music, Pocket Casts, Overcast) and YouTube. In parallel, we expanded into other types of video content for YouTube, including a set of 10 Frontier Math interviews & discussions with contributing mathematicians. In 2026, we plan to continue exploring audiovisual media with more podcasts and videos.

Support our work

We are actively fundraising $3 million, and could productively deploy an additional $10 million in funding to grow our scope.

Here are some concrete short and focused investigations that we are excited about and ready to pursue contingent on securing a funder:

• Inference cost trajectories: investigating whether the cost of inference required to reach a given capability level will continue to fall at the recent rapid pace, or whether improvement rates are likely to slow or plateau.
• Training data trends and bottlenecks: mapping trends in training data supply, composition, and cost, including reinforcement learning environments and synthetic data, and assessing whether data scarcity is likely to become a binding constraint on current scaling trends.
• AI diffusion and adoption: producing a high-resolution picture of AI diffusion, similar to the work in this newsletter post, but using higher-quality datasets and more granular cuts, or running new polls and surveys. We would examine geography (including China and India) and break adoption down further by use case.

Each of these could be a three-month investigation, resulting in a detailed written report, similar in scope and depth to our work with EPRI studying AI power demand. To discuss how you can make one of these projects happen, reach out to donate@epoch.ai.

Larger donations would allow us to launch or expand major new programs such as expanding our satellite tracking of frontier AI data centers to China and other countries, developing new benchmarks tracking frontier capabilities, conducting more consistent evaluations of frontier open models, and mapping AI supply chains to identify key inputs and bottlenecks to AI chip and data center construction.

You can donate through https://epoch.ai/donate, or reach out to donate@epoch.ai to discuss a large grant. Donations of any size are welcome.

Generated by RSStT. The copyright belongs to the original author.

Source