Epoch AI
85 Intel
Epoch AI is a nonprofit research institute that tracks trends in artificial intelligence such as training compute, hardware and model capabilities, and forecasts AI's economic and societal impact. It maintains public databases of machine learning models and created the FrontierMath benchmark.
Intel
Epoch AI estimates how CoWoS packaging, HBM and advanced logic dies were allocated among Nvidia, Google, AMD and Amazon in 2025, and argues packaging and memory were the binding constraints on AI chip production.
An Epoch AI and Ipsos survey of 1,106 US workers finds one in five now delegates tasks to AI instead of coworkers or contractors. AI rarely does a whole task, but saves time most often when it does, and outputs are usually accepted with little editing.
Epoch AI analyzed merged pull requests from 41 Codex contributors with LLM effort estimates. Days showing over 24 hours of equivalent unassisted work grew from 2% to 8%, read as an upper bound on AI uplift.
Epoch AI argues that the ability to divide, coordinate and recombine research work is a missing parameter in intelligence-explosion models, so even fully automated R&D could be bottlenecked and delay a technological singularity.
Epoch measures pairwise correlations across AI benchmarks, finding rankings remain consistent even across domains, with a 0.68 median cross-category correlation.
Epoch AI estimates Huawei's AI compute output at under 4% of Nvidia's in 2026, possibly near 1% by 2028 without foreign memory. It concludes export controls cap Huawei's key scaling levers, so catching up by 2030 is unlikely.
Epoch's index puts GPT-6 Astra at 166 overall and a record 170 in math, while Claude Fable 5.1 keeps the software engineering lead at 167 versus 164.
Epoch AI's early InnovationEval results test whether frontier models can independently invent ML techniques comparable to human research. Despite heavy GPU budgets, models make little progress, a data point on how far end-to-end research agents are from automation.
Epoch AI's dataset of over 500 AI supercomputers shows performance growing 2.5x a year, costs reaching billions, power draw comparable to mid-sized cities, and companies now holding about 80% of capacity as government share declines.
Epoch estimates NVIDIA's B200 costs $5,700 to $7,300 to build, with HBM3E and CoWoS packaging driving most of it, against reported sale prices of $30,000 to $40,000.
Epoch estimates that 290,000 to 1.6 million H100-equivalents reached China illegally through 2025, with a 660,000 median, roughly a third of China's total compute, built from indictments, reporting and a statistical model.
Epoch AI evaluates Gemini 2.5 Deep Think on FrontierMath, mathematician feedback, and IMO problems. It finds better knowledge use and computation, a more conceptual geometry approach, weaker creativity on older IMO problems, and citation issues.
Epoch AI tracks LLM context windows growing about 30x per year since mid-2023, while the input length where top models hit 80% accuracy on long-context benchmarks rose over 250x in nine months.