AI 2030
A · Strong evidence Epoch AI, the main independent tracker of AI compute and investment trends, extrapolates its own datasets to 2030 with sources and benchmark fits shown; commissioned by Google DeepMind and a forecast by design.
Summary based on the full text·Original: about 110 minutes to read
In short
A September 2025 report by Epoch AI, commissioned by Google DeepMind, asking what happens if current AI scaling trends continue to 2030. It argues that compute, investment, data, hardware and energy can all keep growing exponentially, and that benchmark trends then give a baseline for capabilities. Its main application is scientific R&D: desk-based work such as software and mathematics should benefit most, while experimental fields such as drug development will move more slowly because of trials and regulation.
Key numbers
- 1,000xcompute of the largest 2030 training runs relative to today's leading models, on current trends
- 4-5x per yeargrowth in training compute for notable AI models since 2010
- two hundred billion dollarsinvestment for a 2030 frontier training cluster, close to 1% of present US GDP
- 10-20%minimum expected day-to-day productivity gain from AI assistants in non-experimental work tasks
- about 10 GWpower the largest training runs could require by 2030
As reported by the source.
English·PDF·opens epoch.ai
What it reports
- Training compute has grown 4-5x per year since 2010; on trend, 2030 frontier models use 1,000x today's compute.
- A 2030 frontier cluster would cost about two hundred billion dollars; AI lab revenues have grown about 3x per year recently.
- Training datasets grew 2.7x per year; synthetic and multimodal data are expected to cover the shortfall in human text.
- Frontier training power has grown 2.1x per year; the largest runs could need about 10 GW by 2030, and AI data centres could exceed 2% of global electricity demand.
- Benchmarks such as SWE-bench Verified are on track to be solved in 2026; research engineering tasks (RE-Bench) progress more slowly.
- At a minimum, AI assistants should raise day-to-day productivity by 10-20% in non-experimental work, based on Copilot trials from late 2023.
- Drugs approved by 2030 are already in today's pipeline, so AI is unlikely to yield approved products by then under current rules.
How they got it
Trend extrapolation from Epoch's databases of notable models, hardware, costs and power, with stated reasons for possible slowdowns (investor sentiment, regulation, chip or energy supply) or speed-ups (R&D automation). Capability forecasts fit sigmoid curves to benchmark scores, documented in an appendix, and are applied to software engineering, mathematics, molecular biology and weather prediction.
Limits
- Assumes compute drives progress and that trends persist; the authors say it cannot predict general intelligence.
- Benchmarks may not represent the hardest real tasks, and some show no scaling progress yet.
- The productivity floor rests on early Copilot trials, and the authors note negative results exist.
- Funded by a leading AI developer whose investment case depends on continued scaling.
Why it matters for companies
The report gives planners a defensible baseline: assume AI tools keep getting markedly better through 2030 and plan for the gains to land first in digital, desk-based work. Software, analysis and documentation workflows are where capability will outrun organisational readiness. Physical and regulated processes will lag, so business cases that depend on them need longer horizons.
Summary based on the full text. Read the original for the full argument and data.Open PDF on epoch.ai
- Tags
- AI capability