Source
https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf — original source (opens in a new tab; the file is not redistributed)
Context: thesis - No NI feeds (shallow ingest)
Summary
Stanford HAI’s annual AI Index is the reference account of the state of artificial intelligence: research output, technical performance, investment and policy. This page carries its evidence that frontier AI development has concentrated into private industry, behind a compute-cost barrier that grew five orders of magnitude in seven years. Related held concentration evidence: food systems (OT_231), wealth (OT_233) and critical minerals (OT_235).
Key thesis insights
- Frontier AI is made by private industry, not public institutions. “Nearly 90% of notable AI models in 2024 came from industry, up from 60% in 2023” (Top Takeaways, p.4; Chapter 1 highlights restate it as “originating from industry”, p.27), while “academia remains the top source of highly cited research” (p.4). The knowledge is public; the capability is private. OT_236
- The compute barrier compounds rapidly. “Model scale continues to grow rapidly”, with “training compute doubles every five months, datasets every eight, and power use annually” (Top Takeaways, p.4). OT_236
- Training cost rose five orders of magnitude in seven years (Training Cost, p.66,
Epoch AI estimates from cloud compute rental prices): the original Transformer (2017)
“cost around
670 to train"; RoBERTa Large (2019) "around160,000”; GPT-4 (2023) “estimated around79 million"; Llama 3.1-405B (2024) "an estimated cost of170 million”. Frontier model development is open only to organisations commanding hundreds of millions of dollars of compute. OT_236 - Caveat for accurate use: the concentration is structural, not a single leader. The report finds the performance frontier convergent: the Elo gap between the top and 10th-ranked models fell from 11.9% to 5.4% in a year, and the top two are separated by 0.7% (p.4). The citable concentration claim is about who can build frontier models (industry, compute-rich firms), not about one runaway winner. OT_236
Notes
- Related evidence. Together with OT_231 (agri-food supply chains), OT_233 (wealth) and OT_235 (critical-mineral production and refining), this page completes a concentration evidence set spanning food, wealth, minerals and AI/compute.
- Verification. All quoted figures verified verbatim against the held PDF (pdftotext, printed page numbers confirmed off the page footers). sha256 recorded for lint L22.
- Scope. context thesis, feeds empty, shallow ingest. No research target opened.
Connections
Links to