The Epoch AI Brief - November 2025
Frontier Data Center Hub launch, OSWorld analysis, data center build time, "Claudiness" factor behind benchmark scores and our 2025 impact survey.
Hi! In this edition of the Epoch AI brief:
We launched our Frontier Data Centers Hub, mapping the construction of GW-scale AI facilities using satellite imagery and public sources.
We analyzed OSWorld, finding that the benchmark for AI computer use relies on simple tasks and often hinges on interpreting ambiguous instructions.
We published four new Data Insights, covering data center build times, their physical footprint and power usage comparisons.
We released two Gradient Updates, discussing the “Claudiness” factor behind benchmark scores and the need for experiments in the intelligence explosion debate.
We evaluated Gemini 3 Pro, Claude 4.5 Opus & GPT 5.1 on part of our benchmarking suite.
We are hiring for Researcher and Lead Editor roles.
Understanding how our research impacts our audience helps us set priorities in 2026. We are running our annual survey and we want to hear from you.
Publications & Announcements
We launched our Frontier Data Centers Hub
We mapped the construction of the largest US AI data centers using satellite imagery, permits & public sources — releasing everything for free, including commissioned satellite images.
Several data centers will soon demand 1 GW of power, starting early next year, but only a few countries have enough power to build several GW data centers. For example, 30 GW is roughly equivalent to 5% of the US’ average grid load, 2.5% of China’s, but 90% of the UK’s.
AI data centers require massive capital investment, typically $29B per gigawatt of total facility power. By the end of the year, AI data centers could collectively see >$300 billion in investment, around 1% of US GDP. That’s bigger than the Apollo Program (0.8%) and Manhattan Project (0.4%) at their peaks. The most expensive data center we know about is projected to exceed $100 billion in total capital cost upon completion.
Track each data center’s build-out in our Satellite Viewer with annotated high-resolution imagery in our data hub or dive deeper into what’s inside them with our accompanying report “What you need to know about AI data centers”.
OS World review & AI computer use
We looked at OSWorld, a popular evaluation of AI computer use capabilities.
Our findings: Tasks are simple, requiring fewer than ten steps (clicks, text inputs, etc.), taking humans only a few minutes to complete; many don’t require graphical user interfaces (GUIs); and success often hinges on interpreting ambiguous instructions. The benchmark is also not stable over time.
Learn more by reading the full report on our website.
Data Insights
Our Data Insights offer digestible snapshots of complex trends in AI. Since our October newsletter, we’ve published four new insights. Here’s what we’ve found:
Microsoft’s Fairwater datacenter will use more power than Los Angeles
The largest AI data center campuses will soon be a fifth the size of Manhattan
Build times for gigawatt-scale data centers can be 2 years or less
Epoch’s Capabilities Index stitches together benchmarks across a wide range of difficulties
Gradient Updates
Since our October newsletter, we’ve published two new issues of Gradient Updates, our weekly newsletter containing shorter-form research and commentary on important issues on AI:
Benchmarking updates
Gemini 3 Pro set a new record on FrontierMath: 38% on Tiers 1–3 and 19% on Tier 4. It is currently the model with the highest ECI score in our database. Gemini 3 Pro also set a new record on GPQA Diamond: 93% vs. the previous record of 88%. Almost all of this gain came in organic chemistry.
We ran Opus 4.5 across various benchmarks. On FrontierMath, it scored 21% on Tiers 1–3 and 4% on Tier 4. Building on our “Claudiness” Gradient Update, we looked at how Claudey Opus 4.5 is relative to other frontier models.
GPT-5.1 is about as capable as GPT-5, according to our ECI. We see some evidence in favor of Open AI’s claim of “less time on easy tasks and more time on hard tasks”.
See more and explore the date on our benchmarking hub.
Other updates and quick takes
We located and analysed the Stargate UAE data center.
In case you missed the full podcast episode, here is Yafah Edelman’s (Head of Data & Trends at Epoch AI) on why AI adoption may be faster than some people expect.
We updated a chart from a previous Gradient Update on unprecedented revenue growth to include Anthropic’s reported projections in early november.
By stitching benchmarks together, the Epoch Capabilities Index allows us to compare frontier models to models with 100,000x less training compute.
Organizational Updates
Impact Report annual survey
As 2025 closes, we’re gathering stories about how our research has informed decisions, shaped strategies, or sparked new thinking this year. We are looking for specific moments (even tiny ones) where
better data → insight → action
Your feedback helps us double down on what’s working, prune what’s not, and sharpen our 2026 priorities. Take our survey and help us improve!
It shouldn’t take more than 10-15 minutes.
Based on reader feedback, we’ve brought all our content into Substack. See this post on how manage your subscription and receive only the update you care about.