The Epoch Brief - March 2026
AI solves an open research math problem, growth in high-bandwidth chip memory, what job postings reveal about frontier labs' plans, and more
Welcome to this edition of the Epoch AI Brief! This month:
We published three new Data Insights: Microsoft’s AI-driven capex, bottlenecks in AI chip production, and AI chip memory bandwidth growth.
We published two new Gradient Updates: final training runs account for a minority of R&D compute spending, and what job postings reveal about the plans of frontier AI labs.
We saw notable AI achievements on our mathematics benchmarks: a solution to a problem in our FrontierMath: Open Problems, and a new record on FrontierMath: Tiers 1–4.
We’re hiring a Data Scientist (Contract) and a Special Projects Associate.
Publications
Data Insights
We published three new Data Insights, which offer digestible snapshots of complex trends in AI.
Total AI chip memory bandwidth has grown 4.1× per year, now reaching 70 million terabytes per second. As of Q4 2025, the cumulative memory bandwidth of AI chips shipped since 2022 has reached roughly 70 million terabytes per second — around 300,000× more data per second than global internet traffic. Figures are drawn from financial disclosures of five major chip manufacturers.
Advanced packaging (CoWoS) and high-bandwidth memory (HBM), not logic dies, were the bottlenecks on AI chip production in 2025. We estimate that the four largest AI chip designers collectively consumed around 90% of global CoWoS capacity and HBM supply by value in 2025, while consuming only 12% of advanced logic die production.
Microsoft’s recent $68 billion in physical assets additions were driven by AI-related purchases. The company’s financial filings suggest that purchases were dominated by spending on new data centers. IT equipment, including GPUs and servers, contributed 57% of the growth, while buildings made up 39%.
Gradient Updates
We published two new Gradient Updates, our newsletter with shorter-form research and commentary by specific authors on important issues in AI.
Final training runs account for a minority of R&D compute spending. We estimate that less than 30% of R&D compute spending by MiniMax and Z.ai goes to final training runs. This corroborates previous findings for OpenAI from 2024 and suggests a pattern across companies of various sizes and locations.
What do frontier AI companies’ job postings reveal about their plans? AI companies guard their strategies closely, but their hiring pages are public. An analysis of job postings shows a fast increase in go-to-market roles, and hints about upcoming products.
FrontierMath Benchmarks
AI has solved a problem in FrontierMath: Open Problems, our benchmark of real research problems that mathematicians have tried and failed to solve. The problem is from the Moderately Interesting tier. The contributing mathematician, Will Brian, is excited to write up the solution for publication, along with Kevin Barreto and Liam Price, who first elicited the solution from GPT-5.4 Pro.
We removed a different problem from FrontierMath: Open Problems after determining that it did not meet our bar for mathematical notability.
Earlier in March, GPT-5.4 set a new record on FrontierMath: Tiers 1–4, our benchmark of extremely challenging math problems with known solutions. 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%, including a solution to one Tier 4 problem that no model had solved before.
Other quick takes and updates
We updated the epoch.ai website design. Our goal is to 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 to hear what you think — email us at design@epoch.ai
We’ve updated the Epoch Capabilities Index with three new benchmarks (APEX-Agents, ARC-AGI-2, and HLE) and several new models. GPT-5.4 Pro is the current state-of-the-art, with a narrow lead over Gemini 3.1 Pro.
Organizational Updates
Careers
We’re hiring for a Data Scientist (Contract) to assist with our AI research efforts. This role involves reviewing technical literature, tracking benchmark data, compiling technical infrastructure details, and analyzing various sources to build comprehensive insights about AI models, data centers, and companies.
We’re also looking for a Special Projects Associate to provide operational support for new projects, including creating new benchmarks.
Applications are rolling, so apply soon! As always, all roles are fully remote.
Based on reader feedback, we’ve brought all our content into Substack. See this post on how to manage your subscription and receive only the updates you care about.