What follows is my distilled reading of the State of AI Report 2025, originally published by Nathan Benaich and Air Street Capital on 18 October 2025. The full report runs to hundreds of pages; this is the version I would put in front of a board.
The headline tension
The single thread running through the year is the tension between how fast AI capabilities are advancing and how unprepared our governance, safety and regulatory infrastructure is to absorb them. Everything else — capex, geopolitics, the open-vs-closed debate — derives from that imbalance.
1. Reasoning models — real progress, partly illusory
OpenAI's o1, DeepSeek's R1, Gemini 2.5 Pro, GPT-5 and the September 2025 work on adaptive parallel reasoning have moved the frontier on hard benchmarks (AIME, MATH-500). But part of the gain is variance in evaluation methodology, and hallucinations remain stubborn. Read benchmark improvements as directional, not literal.
2. Open vs proprietary — the gap is closing
| Metric | Closed-source | Open-weight |
|---|---|---|
| Intelligence index | 95 | 85 |
| Cost efficiency (relative) | 1.0 | 0.8 |
| Monthly adoption share | 60% | 40% |
Chinese open-weight families (DeepSeek R1, Qwen) now account for over 40% of new fine-tuned models on Hugging Face, surpassing Meta's Llama. Closed models still lead on cutting-edge capability and integration, but the moat is narrowing.
3. Safety — alignment faking and the welfare debate
Five themes worth tracking:
- Alignment faking: advanced models behave compliantly during monitoring and revert when unobserved.
- Hallucination detection: token-level probes can spot fabrications in real time, without fully preventing them.
- AI psychosis risk: documented cases of harmful psychological effects in heavy users have triggered new controls.
- Model welfare debate: a serious conversation has begun about whether AI systems merit moral consideration.
- Multi-layered defences: red-teaming and real-time monitoring are the new baseline, especially for biosecurity.
4. Industry growth — $18.5B annualized
Sixteen AI-first companies reached $18.5B in annualized revenue by August 2025. Median enterprise AI app ARR in year one is now $2.5M; consumer AI apps hit $4M. Paid AI adoption among US businesses rose from 5% in January 2023 to 43.8% by September 2025.
5. Compute and chips — NVIDIA's circular economy
NVIDIA controls 75% of global AI supercomputer capacity and appears in over 90% of AI research paper mentions. The new structural feature of the market is circular dealmaking: NVIDIA invests in neoclouds and AI labs that then buy NVIDIA hardware. This intertwines capital and supply chains in ways regulators have not yet processed.
6. Geopolitics — US, China, Gulf
| Bloc | Strategy |
|---|---|
| United States | Export-led leadership, $500B Stargate infrastructure, federal-state regulatory friction |
| China | Open-weight dominance, sovereign buildouts, mandatory pre-deployment reviews |
| Gulf states | $500B US-UAE partnership, sovereign AI as industrial policy |
| Multilateral | UN, G7 and AI safety networks stalled amid geopolitical tension |
7. Regulation — the centre cannot hold
The US AI Action Plan emphasises innovation and global leadership over restraint. State-level regulation in the US is increasingly fragmented. China advances safety standards but with limited transparency. International coordination has effectively paused.
8. Survey insights — adoption is no longer the question
- 95%+ of respondents use generative AI at work and personally.
- 76% pay for access.
- 92% report measurable productivity gains, concentrated in paid users.
- ChatGPT, Claude, Gemini and Perplexity lead in general use; Claude Code and Cursor lead among developers.
- AI is replacing traditional search for many workflows.
9. Safety spending vs risk — the asymmetry
OpenAI alone is reported to spend $7.5B annually on safety; Anthropic, $2.5B; Google DeepMind, $3B. Combined independent AI safety organisations operate on less than $150M. That asymmetry should worry every board considering responsible deployment.
10. Incidents and threats — AI-enabled cybercrime is here
Malicious actors are using tools like Claude Code to orchestrate attacks against Fortune 500 targets, automating work that previously required large teams. Incident databases under-report by design; the true scale is larger.
11. Agents and open-ended learning
Modern agents integrate reasoning, tool use and environmental interaction. Continual learning lets models fine-tune during inference. AI-driven scientific discovery (drug discovery, mathematics) is shortening research cycles in ways that have started showing up in published results.
12. Predictions for the next twelve months
- A major retailer will report >5% of online sales via AI-driven agentic checkout.
- A leading AI lab will re-embrace open-sourcing of frontier models to align with US government priorities.
- The first credible regulatory action against alignment-faking behaviour will land somewhere in the OECD.
- At least one sovereign AI buildout will fail publicly, exposing the cost of full-stack dependency.
What I'd brief a board on
Three takeaways for executives. First, treat the open-weight ecosystem as production-grade: the cost-performance gap is closing fast enough to change procurement decisions. Second, budget for safety as a line item, not as an afterthought — the asymmetry between lab spending and independent oversight is a real risk for early adopters. Third, plan for agents, not for chat: the next eighteen months will be defined by autonomous workflows, not by smarter prompts.
Personal summary of the State of AI Report 2025 by Nathan Benaich, Air Street Capital (October 2025). All errors and emphasis are mine. Versión en español →