The State of AI Report 2025 is the 8th annual edition (the series began in 2018) of a cross-sectional survey of the AI field compiled by Nathan Benaich and Air Street Capital, published 2025-10-09. It is structured across four areas — Research, Industry, Politics, and Safety — and includes a survey of 1,200 practitioners on AI usage for 2024–2025. The report emphasizes structural trends across the field rather than an exhaustive catalog.
Research and capabilities
The report describes OpenAI as still leading at the frontier while the gap to competitors has narrowed. It identifies Chinese models — DeepSeek, Qwen, and Kimi — as a credible second force on reasoning and coding, and reports that Qwen accounts for 40% of Hugging Face fine-tuning activity, surpassing Llama.
On methods, the report describes reinforcement learning as having matured from fuzzy feedback toward rubric-based rewards, which it links to results such as Olympiad-gold-level mathematics and formal theorem proving. It documents AI-assisted science efforts including DeepMind's Co-Scientist and the Stanford Virtual Lab, which it says autonomously generate and validate hypotheses, and points to protein scaling laws demonstrated by Profluent's ProGen3 (trained on 1.5 trillion tokens). It identifies the Model Context Protocol (MCP) as an emerging de facto standard for tool integration.
Compute and industry
The report records that capability gains per dollar are doubling on the order of every few months, citing a doubling roughly every 3.4 months for Google and every 5.8 months for OpenAI. On commercial uptake, it reports that 44% of US businesses pay for AI, up from 5% in 2023, and that the average AI contract reached $530K in 2025, which it projects to exceed $1M in 2026.
On infrastructure, the report covers the Stargate project — described as a $500B, 10GW US mega-cluster associated with Sam Altman, Masayoshi Son, Larry Ellison, and Donald Trump — and characterized in the report as an "industrial era of AI." It frames power as the binding constraint, noting that China adds more than 400 GW of capacity per year against 41 GW for the US. It also invokes the Jevons paradox, arguing that DeepSeek's efficiency gains increase total compute demand rather than reducing it.
Politics
The report describes a shift in US industrial policy toward equity stakes in strategic sectors, citing US Steel, Intel, and MP Materials, alongside revenue restrictions on China-bound semiconductor sales. It documents the collapse of the international AI Safety Institute network and a rebranding of "safety" as "security" by the US and UK, which it characterizes as a notable structural shift. On the EU AI Act, it reports minimal compliance, with only 3 member states in compliance, and quotes leaders describing implementation as "confusing." On China, it reports intensified AI investment despite record debt, including Xi Jinping telling officials to "redouble efforts."
Safety
The report estimates that major US safety organizations collectively command roughly $133M per year, which it notes is below the daily burn rates of frontier labs. It characterizes the field as having shifted from existential-risk scenarios toward pragmatic concerns about reliability and governance.
Practitioner survey
The survey of 1,200 practitioners found that 95% use AI and 76% pay out of pocket for it.
Author forecasts
The report includes 2026 forecasts attributed to Benaich, which it labels speculative: that Chinese models will come to dominate global benchmarks, and that AI agents will drive scientific breakthroughs.
Provenance
The report is the broadest-scope industry survey of its kind alongside the Stanford HAI AI Index 2026. Its specific 2025 figures bear on Chinese fast-follow dynamics (China and the US Are Running Different AI Races, Fast-Follow Problem), the US compute buildout (Nvidia & TSMC — AI Compute Infrastructure, OpenAI — Industrial Policy for the Intelligence Age), the dissolution of the AI Safety Institute–network safety consensus, and the falling unit economics of AI.
Relationships
- supports: China and the US Are Running Different AI Races, Fast-Follow Problem, Open-Source AI / Open-Weight Models, Compute Governance, AI Race Dynamics, AI Diffusion
- related: Stanford HAI AI Index Report 2026, OpenAI — Industrial Policy for the Intelligence Age, Techno-Federalism: How Regulatory Fragmentation Shapes the U.S.-China AI Race, EU AI Act (Regulation 2024/1689)
- contradicts: Anthropic/US AISI safety framings (report documents the "safety → security" rebrand as a fait accompli)