Every AI chip consists of key components. If the supply of any component stalls, so can chip production. Introducing our AI Chip Components explorer, which tracks supply and consumption of logic, packaging, and memory to identify the bottlenecks shaping frontier AI compute.

We're looking for a new Designer to join our team! The Designer will help design dashboards and data visualizations that make our research more accessible and engaging.

We are launching domain-specific capability scores, tracking the capabilities of models across SWE and Math benchmarks, using the same scale as the general ECI. We also support customization for users who want to create their own variants of the ECI. Link below!

The recipe for “classic” reasoning benchmarks is simple: text-only, several-hour time horizons, easy to grade, with expert human baselines. What next? In this week’s Gradient Update, Greg Burnham argues it’s as easy as dropping one of these four ingredients.

Join us for in-person workshops to develop problems for FrontierMath: Open Problems! We are seeking highly interesting unsolved problems from research mathematics whose solutions can be verified programmatically. These are hard to find. Come take a crack at it! Link below.

How much AI compute has been smuggled to China? We estimate between 290k and 1.6M H100-equivalents by the end of 2025 — representing ~20% to ~60% of China’s total compute.

GPT-5.5 Pro achieves a new high score of 159 on the Epoch Capabilities Index! ECI is our statistical tool that combines multiple benchmarks into a unified scale.

Alongside our report on how fast robot production could scale up, we've also released a series of @metaculus forecasting questions on the future of robotics. If you have views on where these trends are headed, come weigh in!

How fast could production of humanoids, quadrupeds, drones, and other robots scale up, in the event of a large demand shock?

We now estimate that only about 0.3 GW of total facility power is operational for Stargate Abilene, not 0.6 GW. We've moved the 0.6 GW milestone to late May and the 1.2 GW milestone from Q3 to Q4 2026, but both are uncertain. More about this change and our methodology in 🧵 bsky.app/profile/epo...

Yesterday, we shared a chart showing 80% of Claude users live in $100k+ households, more than any other major AI service. But Claude’s user base is smaller than other AI services, so this isn’t the same as being the most popular service among high-income households. bsky.app/profile/epo...

80% of US adults who report using Claude in the previous week live in households earning $100,000 or more a year, compared to 37% of Meta AI users. Other major providers cluster in a relatively narrow band, with 56–64% of users in $100,000+ households.

Opus 4.7 scores 156 on ECI, our tool for combining multiple benchmarks onto a single scale. This puts it a bit ahead of Opus 4.6 and a bit behind only GPT-5.4, Gemini 3.1 Pro, and GPT-5.4 Pro. Thread with individual scores and commentary.

In 2025, OpenAI announced Stargate, a $500 billion data center initiative. We surveyed all 7 US sites and found visible development at each. There's a long road ahead, but the project appears on track to reach 9+ GW by 2029—comparable to New York City's peak power demand. 🧵

Have AI capabilities accelerated? On 3 out of the 4 AI capability metrics we investigated, we found strong evidence of acceleration, around when reasoning models emerged.

According to our latest polls, Claude usage in the US rose by over 40% amid increased attention last month, but remains far behind ChatGPT. Our point estimate would imply several million new weekly users in the United States.

Errata: Yesterday, we discovered that some of our chip owner estimates were stale—Oracle's Nvidia compute wasn’t subtracted from "Other" as intended. This inflated “Other” by ~1M H100e, 5% of the overall total. In our corrected figures, hyperscalers hold 71% of world AI compute.

OpenAI has purchased access to the FrontierMath: Open Problems verifiers. This allows them to check the validity of solutions their models generate. Thread with details.

@NBCNews on our recent AI usage survey: www.nbcnews.com/tech/tech-n...

Five companies — Google, Microsoft, Meta, Amazon, and Oracle — now control about two-thirds of the world's compute, up slightly from ~60% at the start of 2024. Many AI labs (including OpenAI and Anthropic) depend almost entirely on these hyperscalers for access to their compute.

What are the largest software engineering tasks AI can perform? In our new benchmark, MirrorCode, Claude Opus 4.6 reimplemented a 16,000-line bioinformatics toolkit — a task we believe would take a human engineer weeks. Co-developed with @METR_Evals. Details in thread.

The Iran War and Hormuz shutdown have disrupted oil, gas, and helium exports and threatened data centers and investments in the Gulf states. @justjoshinyou.bsky.social explores how a prolonged Iran war could affect AI, and why it probably won’t completely derail the compute buildout.

AI is changing our jobs: among people who use AI regularly at work, 27% say AI has replaced some of their tasks; 21% say it has enabled new tasks. This and other workplace usage findings from our new Epoch AI/Ipsos survey on AI usage in 🧵

We had pre-release access to Meta’s new Muse Spark model and evaluated it on FrontierMath. It scored 39% on Tiers 1-3 and 15% on Tier 4. This is competitive with several recent frontier models, though behind GPT-5.4.

New essay by Anson Ho: Chinese and open model AI labs have ≈10× less compute than the frontier. But they can distill frontier models, replicate innovations fast, and have enormous talent. Is that enough to compete at the frontier? 🧵

Who owns the world's compute? Our new Chip Ownership hub shows that Google leads, holding around 25% of all compute sold since 2022.

Compute may be the most important input to AI. So who owns the world’s AI compute? Introducing our new AI Chip Owners explorer, showing our analysis of how leading AI chips are distributed among hyperscalers and other major players, broken down by chip type over time.

We have removed a problem from FrontierMath: Open Problems. The problem was solved by AI, but upon review we determined that the problem didn’t meet our minimum bar for mathematical notability. This is a different problem from the one whose solution we announced on Monday.

The total memory bandwidth of AI chips shipped since 2022 has reached 70 million terabytes per second, growing 4.1x per year. That's around 300,000x more data per second than global internet traffic.

AI companies guard their strategies closely. Their hiring pages, however, are public. JS Denain and Campbell Hutcheson analyzed open roles at OpenAI, Anthropic, xAI, and DeepMind. Here's some of what their job postings tell us about their plans:

How do AI companies allocate their R&D compute? JS Denain and Cheryl Wu estimate that across OpenAI, MiniMax, and Z.ai, less than 30% of R&D compute spending goes to final training runs.

AI has solved one of the problems in FrontierMath: Open Problems, our benchmark of real research problems that mathematicians have tried and failed to solve. See thread for more.

How much of the world's advanced chip packaging and high-bandwidth memory does AI consume? Almost all of it. We estimate the four largest AI chip designers consumed ~90% of global advanced packaging and HBM supply in 2025, suggesting these inputs were bottlenecks in 2025.

We've updated the Epoch Capabilities Index with three new benchmarks (APEX-Agents, ARC-AGI-2, and HLE) and several new models! In our latest estimate, GPT-5.4 Pro sets a new SOTA at 158, a narrow lead over Gemini 3.1 Pro's score of 157.

We've updated epoch.ai Our goal: make it easier to find and use our work, whether you're a policymaker, researcher, or just curious about where AI is headed. We'd love your feedback. Comment below or reach out to [email protected].

Microsoft added $68B in physical assets in the second half of 2025 — almost as much as the entire prior fiscal year. 57% was IT equipment (GPUs, servers). 39% was buildings, dominated by data centers.

GPT-5.4 set a new record on FrontierMath, our benchmark of extremely challenging math problems! We had pre-release access to evaluate the model. On Tiers 1–3, GPT-5.4 Pro scored 50%. On Tier 4 it scored 38%. See thread for commentary and additional experiments.

OpenAI’s recent funding round nearly triples the amount they have raised so far. The Information has reported that OpenAI projects a $157B cash burn through 2028. This round, combined with $40B cash on hand, essentially matches that projection.

Our team is expanding! We’re hiring for a Researcher to work with our Senior Researcher, JS Denain, on pitching and developing new projects, and a Special Projects Associate to provide operational support for making great new benchmarks.

Driven by investments in AI, hyperscaler capital expenditures have grown 70% per year since the release of GPT-4, nearing half a trillion dollars in total during 2025. If this trend continues, Alphabet, Amazon, Meta, Microsoft and Oracle will spend $770 billion on capex in 2026.

Developing more powerful AI isn’t just about scaling compute. It’s also about improving algorithms and data quality, which let you build better models with the same compute. We call this “AI software progress” — here’s everything you need to know about it: 🧵

In 2024, Epoch AI estimated the rate of software progress in language models. We found that training compute efficiency was improving at ~3x per year. But this estimate was for pre-training, and is now outdated — so Epoch AI researcher Anson Ho took a new look at the numbers. 🧵

AI training compute efficiency has improved extremely fast: each year, you need several times less training compute to reach the same capability. But AI architectures/algorithms haven’t changed *that* much in recent years. So where do these efficiency improvements come from? 🧵

AI is getting better at math almost as fast as we can write new benchmarks to test it. IEEE Spectrum just published a piece featuring our past & present work on FrontierMath — as well as Aletheia and First Proof. Epoch researcher Greg Burnham called this “a more-the-merrier situation.”

Commissioned by Sentinel Bio, we have released a database of over 1,100 AI models in biology across nine categories, from protein engineering to pathogen property prediction. We analyzed their safeguards, accessibility, training data, and the foundation models they build on.

Last year, we noticed a gap between scores in our SWE-bench Verified runs and scores reported elsewhere. We’ve now updated our evaluation methodology. For most models, we’re seeing scores close to those reported by the original model developers.

Gemini 3.1 Pro scored comparably to Gemini 3 Pro on FrontierMath. It also solved a Tier 4 problem that no model has solved before, though not how a human would. See thread for the problem author’s reaction.

OpenAI may be a household name, but Anthropic could soon be earning more revenue. Since each company hit $1B in annualized revenues, Anthropic has grown substantially faster (10× vs 3.4× per year) and could overtake OpenAI by mid-2026 if recent trends continue.

Toby Ord argues that RL scaling for AI mainly works by allowing models to work for longer, so the gains come with large, per-use inference costs. According to JS Denain, this overstates how much of a bottleneck this will be for AI progress.

Can AI do real digital work? We reviewed three benchmarks to find out: RLI, GDPval, and APEX-Agents. Our take: progress here will indicate substantial economic value, but tasks are too self-contained to tell us about wholesale automation. Thread for more: