Epoch AI

Epoch AI estimates the effective stock of public human-generated text at roughly 300 trillion tokens and projects that language models will fully use it between 2026 and 2032, sooner if heavily overtrained.

Epoch AI reviews three benchmarks of economically valuable digital work, comparing task sourcing, realism, human time estimates and evaluation methods. Progress should track utility, but the self-contained tasks fall short of indicating full job automation.

Epoch AI compares China's recorded $3.8B of Malaysian server imports with Malaysia's $0.6B recorded exports from April 2024 to June 2025, a pattern consistent with about 150,000 H100-equivalents of smuggled compute, without proving diversion.

Epoch AI's overview of the AI data center buildout: scale, power demand, what distinguishes these facilities from conventional ones, and what the buildout implies for AI progress and policy.

Epoch AI reports that models fitting on a single top consumer GPU such as the RTX 5090 match frontier LLM performance from 6 to 12 months earlier, while warning that benchmark tuning of small models may understate the real gap.

A 2026 Epoch AI and Ipsos survey of over 2,000 US adults finds half of employed AI users rely on it at least as much for work as for personal tasks. AI has replaced tasks for 27% of work users and created new ones for 21%.

Epoch AI estimates training costs for large-scale ML models, amortizing cluster costs over training time, and finds spending doubling roughly every eight months. About half goes to GPUs, the rest to other hardware and energy.

Epoch's domain-specific capability indices show Claude models consistently beating their general score on software engineering while trailing on math, though Opus 4.6 and 4.7 narrow the gap.

Epoch AI shows the record for single AI data center IT power capacity has doubled every ten months since mid-2024, with xAI's Colossus 2 estimated at 950 MW at the current top.

Epoch AI data insight on the physical footprint of frontier AI data center campuses, with power, H100-equivalent compute and cost for sites like Meta Hyperion, xAI Colossus 1 and Stargate Abilene, compared against Manhattan and Central Park.

Epoch AI compares ML hardware performance by number format. Tensor-FP16, TF32 and tensor-INT8 give roughly 6x, 11x and 13x the FP32 average, with some chips such as the H100 reaching 59x in INT8.

Epoch AI dissects OSWorld: many tasks can be done through terminal or Python, instructions are often ambiguous, roughly 10% of tasks have serious errors and 10% depend on live web data, and instructions change over time, limiting through-time comparisons.

Epoch derives agent capacity from high-bandwidth memory shipments and serving benchmarks. Chips shipped through 2027 could run tens to hundreds of millions of frontier agents, or billions of cheaper-model agents, with code and data provided.

Epoch AI tracks monthly high and critical CVEs from 21 notable organizations. July 2026 hit about 2,500, five times the pre-Mythos monthly record and 60% above June, following Anthropic's Project Glasswing.

Epoch AI and METR's MirrorCode benchmark shows agents reimplementing real software from specs. Claude Opus 4.6 rebuilt a roughly 16,000-line, 40-command bioinformatics toolkit, a task estimated at 2 to 17 human weeks. Results are preliminary and raw data is public.

Epoch AI tracks the largest AI data center by computing capacity, finding the record has doubled every seven months since August 2024, with the lead passing between Anthropic-Amazon New Carlisle, Microsoft Fairwater and Meta Prometheus.

Epoch AI measures the cost of reaching a fixed benchmark performance falling about 47% per quarter, or 13x a year, since 2023. That is faster than DNA sequencing, compute, or batteries, with math falling faster than game-based puzzles.

Epoch AI itemizes the total cost of ownership of a one-gigawatt AI data center, from servers and facility to network and energy, annualizing CapEx by asset lifespan. Servers dominate at roughly $5B of $8.5B per year.

Epoch AI finds every frontier model since 2023 came from the US, with Chinese models trailing by an average of seven months (min 4, max 14) on its Epoch Capabilities Index.

Epoch AI and Ipsos polls from June and September 2026 find no detectable increase in the share of US adults reporting cyber incidents since Claude Fable 5 launched, with confidence intervals by incident type.

Epoch AI measures accuracy and runtime across major LLM API providers on GPQA Diamond, MATH Level 5 and OTIS Mock AIME, finding a Pareto frontier where cutting error in half generally slows responses by roughly 2x to 6x.

Epoch reports GPT-6 Astra hit perfect scores on over half of EBR-bench attempts through one overpowered card, which is now banned, and finds multi-agent scaffolds may aid exploration but do not improve results much.

Epoch AI estimates global AI data center capacity from rated chip power and a facility multiplier, showing about 30 GW by Q4 2025 and quarterly additions around 2 GW. Useful grounding for compute supply arguments.

Epoch finds quarterly AI chip purchases deliver 49% more performance per dollar each year, flat through mid-2024 then jumping as Blackwell chips reached most spending.