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 But this estimate was for pre-training, and is now outdated — so Epoch AI researcher Anson Ho took a new look at the numbers.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:

How far are autonomous robots from real-world deployment? We reviewed robot capabilities across 14 concrete tasks. The pattern: navigation is deployed at scale, manipulation mostly is not.

How can we anticipate when AI will be able to do our jobs? You could try building benchmarks as a leading indicator of automation, but they fail to capture the complexity of real-world tasks. So Epoch AI researcher Anson Ho argues for an alternative: try automating tasks in your own job. 🧵

Opus 4.6 did well on FrontierMath. Its score of 40% on Tiers 1-3 is statistically tied with the previous top score, GPT-5.2 (xhigh)'s 41%. This is the first time an Anthropic model has been on the frontier of this benchmark.

Even with some of the highest pay in tech, AI companies' biggest cost seems to be compute, not staff. Across three companies where we have data (Anthropic, Minimax and Z.ai), compute accounts for more than salaries, marketing, and all other spending combined.

Kimi K2.5 set a new record among open-weight models on the Epoch Capabilities Index (ECI), which combines multiple benchmarks onto a single scale. Its score of 147 is about on par with o3, Grok 4, and Sonnet 4.5. It still lags the overall frontier.

Shovel-ready short investigations seek funding! - How is AI’s adoption varying across roles, sectors, & regions? - Trends & bottlenecks for data supply? (incl. synthetic and RL environments) - Forecast for worldwide compute buildout? - Will inference costs continue falling?

How does math research change when the cost of trying your first dumb idea goes to zero? University of Toronto mathematician Daniel Litt joins hosts Greg Burnham & Anson Ho to discuss what today’s models can and can’t do in math, and how far they are from doing high-quality research. Video below!

Was serving GPT-5 profitable? According to jsevillamol.bsky.social, @exponentialview.skystack.xyz’s Hannah Petrovic, and Anson Ho, it depends. Gross margins were around 45%, making inference look profitable. But after accounting for the cost of operations, OpenAI likely incurred a loss.👇

Can AI solve math research problems that have eluded human mathematicians? Our new benchmark, FrontierMath: Open Problems, is designed to help find out. AI hasn’t solved any of these yet, but the game is young!

We’ve added new trends & figures categories to our Trends page! Do you know: • How fast LLM inference prices are falling? • How fast compute stocks are growing? • How long it takes to build a GW scale data center? Find out on our Trends page: epoch.ai/trends

Models that are good at math benchmarks tend to be good at coding and reasoning benchmarks too, pointing to a common factor driving AI capabilities. We find that AI benchmark scores are nearly as correlated across domains (0.68) as within them (0.79).

New record on FrontierMath Tier 4! GPT-5.2 Pro scored 31%, a substantial jump over the previous high score of 19%. Read on for details, including comments from mathematicians.

xAI's Colossus 2 data center is running, but likely won't reach 1 GW of power until May, despite prior claims by Elon Musk. Our updated analysis shows the facility lacks the cooling capacity to run 550,000 Blackwell GPUs at full power, even in winter conditions.

AI data centers can now use as much power as New York State uses on the hottest days of the years. We find that data centers currently have a total capacity of around 30 GW.

How well did forecasters predict 2025 AI progress? According to the AI Digest's survey, forecasters: - Mostly nailed benchmark scores - Underestimated risks from AI-enabled bioweapons - Underestimated revenue by almost 2× - Overestimated public concern about AI Details in 🧵

Our 2025 Impact Report is out. The AI industry is scaling exponentially - investment, compute, data center buildouts. So, it turns out, is demand for making sense of it all. See how we’ve kept up!

We loved this quick visual rundown of the Frontier Datacenters hub. Thanks to Rowan Cheung for featuring our project! www.youtube.com/shorts/szAW...

Frontier labs are investing massively in RL environments, yet most of what happens in this space stays behind closed doors. @chrisbarber and @js_denain interviewed 18 people from RL environment startups, neolabs, and frontier labs. Here's what they found:

Anthropic's data center in Indiana is likely the largest in the world today: 750 megawatts by our calculations. Soon, it will pass the gigawatt milestone. How did they do it, and why do we think it's this big? 🧵

Total AI compute is doubling every 7 months. We tracked quarterly production of AI accelerators across all major chip designers. Since 2022, total compute has grown ~3.3x per year, enabling increasingly larger-scale model development and adoption. 🧵

Global AI compute capacity now totals over 15 million H100-equivalents. Our new AI Chip Sales data explorer tracks where this compute comes from across Nvidia, Google, Amazon, AMD, and Huawei, making it the most comprehensive public dataset available.

Since 2023, every model at the frontier of AI capabilities has come from the United States. Chinese models have trailed by 7 months on average.

It's been a big year for Epoch – we published more plots in 2025 than in all previous years combined!

The scale of decentralized AI training runs over the internet has grown exceptionally fast in recent years. Could they catch up to the frontier of compute? 🧵

Benchmarking is crucial to keep track of AI progress. However, benchmarking is hard: each step of the pipeline involves moving parts that can affect the final headline result. Let's dive deeper:

AI capabilities accelerated in 2024! According to our Epoch Capabilities Index, frontier model improvement nearly doubled, from ~8 points/year to ~15 points/year.

Looking back on 2025: what were our most popular short-form research posts? We published 36 Data Insights and 37 Gradient Updates newsletters. Here's what our website readers found most interesting: 🧵

We benchmarked several open-weight Chinese models on FrontierMath. Their top scores on Tiers 1-3 lag the overall frontier by about seven months.

We took an updated look at the evidence on AI adoption. Key finding: AI adoption has continued to grow faster than almost any other technology in history — but the drivers of this have started to change. 🧵

Over the past month we've partnered with @blueroseorg to survey 5,660 Americans about their AI usage habits. Today we're releasing the results. A key takeaway: A majority of Americans use AI on a weekly basis, with 35% using ChatGPT, 24% Gemini, and 13% Meta AI.

Gemini 3 Flash scored 36% on FrontierMath Tiers 1–3, comparable to top models. It scored comparatively less well on the harder Tier 4.

GPUs only account for about 40% of total power in a frontier AI data center at peak operation. Where does the rest go? 🧵