Epoch AI

Epoch AI breaks down LLM accuracy on AIME and USAMO/IMO problems by difficulty: models near saturate easier tiers and reach IMO gold-equivalent totals, yet none has solved a problem in the hardest tier, with small samples noted.

Epoch AI estimates amortized hardware and energy costs for up to 45 frontier models, finding training cost grows about 2.4x per year and could pass a billion dollars by 2027. Source code and datasets are public.

Epoch AI shows combined capital spending at Alphabet, Amazon, Meta, Microsoft and Oracle growing 72% per year since mid-2023, near $500 billion in 2025, and projects $770 billion in 2026 if the trend continues.

Epoch examines four limits on AI training scale: power, chip manufacturing, data and latency. It projects that training runs near 2e29 FLOP will be feasible by 2030 and discusses which constraint is most limiting.

Epoch AI models inference speed against cost, finding latency scales with the square root of model size and the cube root of memory bandwidth. It argues faster serving is increasingly valuable for reasoning-heavy and agentic workloads, and releases source code.

Epoch AI estimates total NVIDIA compute has grown about 2.3x per year since 2019, doubling every 10 months. Hopper-generation chips make up 77% of the stock, and older generations fall below half of cumulative compute within about four years.

Epoch AI presents new empirical techniques and estimates for the returns to software R&D, and asks whether those returns imply hyperbolic, singularity-like growth in AI progress driven by algorithmic improvements.

Epoch AI estimates Amazon, Google, Meta, Microsoft and Oracle held 71% of cumulative AI compute in H100-equivalents at the end of 2025, up from 63% in Q1 2024, with per-company quarterly breakdowns.

Epoch AI and EPRI project that the largest frontier training runs will draw 4-16 GW by 2030, extrapolating from power requirements that have more than doubled each year, and discuss what that means for the energy sector.

Epoch AI's paper fits a sigmoid model relating model capability to benchmark difficulty, stitching many benchmarks into one scale. It estimates algorithmic progress and projects capability trends. Notebooks are released.

Epoch AI analyzes OpenAI's published usage data: researchers' coding-agent spend, valued at API prices, doubled roughly monthly in 2026. Median reached $601 per day and the 90th percentile exceeded $7,000.

Epoch AI measures the open-weight lag on its Capabilities Index at roughly three months on average, an ECI gap of about 7 points, comparable to the step from o3 to GPT-5.

Epoch AI examines five techniques that trade training compute for inference compute, estimating that 1-2 orders of magnitude more inference can offset roughly one order of magnitude of training, and draws governance implications.

Epoch AI and Ipsos polling finds US adults using AI at least six days a week rose from 8% in March 2026 to 19% in August, while single-day users fell from 17% to 10%, based on samples of about 2,000 and 1,000 respondents.

Epoch AI works through failure rates for 100,000 to million-GPU clusters and argues GPU-memory checkpointing keeps training viable at that scale. It concludes hardware failures will not be the binding limit on frontier training runs.

Epoch AI's chip production dataset puts total AI computing capacity growth at roughly 3.3x per year since 2022, with Nvidia supplying over 60% and Google and Amazon most of the rest.

Epoch AI tracks all seven US Stargate sites: Abilene already runs about 0.3 GW, and the full set plans over 9 GW by 2029, roughly 20 million H100-equivalents. A grounded look at where frontier compute capacity is actually coming online.

Epoch AI measures how fast LLM inference prices fall at fixed performance levels across six benchmarks: GPT-4-level PhD science question performance dropped 40x per year, with rates ranging 9x to 900x depending on the milestone.

Epoch AI projects what AI scaling to 2030 would entail: hundreds of billions in investment, gigawatts of power, and capabilities that could speed scientific R&D. Commissioned by Google DeepMind; conclusions are the authors' own.

Epoch AI reads Microsoft's filings to show $68B of physical asset additions in the second half of 2025, nearly equal to the prior full year, with servers and network equipment at 57% of growth and buildings at 39%.

Epoch AI finds math preprints on arXiv acknowledging AI use rose from 4% in April 2026 to 25% in August, with about 6% crediting substantial research contribution, and the trend holds among authors who published before 2023.

Epoch AI replicates the parametric scaling law from the Chinchilla paper, identifies issues in the original estimates, and presents fits that match the data better and agree with the authors' other methods, informing compute-optimal model and token sizing.

Epoch AI measures the gap between top open-weight and closed models on its Capabilities Index. Since January 2026 the average gap is four months, or 8 ECI points, comparable to the distance between GPT-5 and GPT-5.5.

Epoch AI finds benchmark response lengths rising: reasoning models grow about 5x per year versus 2.2x for non-reasoning models, and currently use roughly 8x more output tokens per question.