🪩 The State of AI Report 2025 🪩
Hi everyone!
The day is finally here: I’m thrilled to share the State of AI Report 2025 with you!
In short, it’s been a monumental 12 months for AI. Our eighth annual report is the most comprehensive it’s ever been, covering what you need to know about research, industry, politics, and safety - along with our first State of AI Usage Survey of 1,200 practitioners. The State of AI Report has become the most widely read and trusted report on AI progress and is our open-access contribution to the AI ecosystem.
Below, I share my director’s cut - a snapshot of key themes and ideas that stood out to me.
I’d appreciate your help in spreading the report far and wide - thanks in advance! Any comments, critiques, or suggestions, please hit reply :-)
In true State of AI Report style, let’s dive into the Research section. 12 months on, OpenAI still leads, but the pack has closed in fast. China’s DeepSeek, Qwen, and Kimi sit within a few points on reasoning and coding. While the US holds the frontier, China is now a credible and popular #2.
Once a “Llama rip-off,” Qwen now powers 40% of all new fine-tunes on Hugging Face. China’s open-weights ecosystem has overtaken Meta’s, with Llama riding off into the sunset…for now.
Reinforcement learning has grown up. After fuzzy human feedback came rubric-based rewards and verifiable reasoning tasks. We’re rediscovering rigor and environments for agents to undertake long-running tasks is all the rage.
What’s this approach enabling? OpenAI and Gemini models both hit math Olympiad gold. Open provers like Gödel-LM are publishing formal proofs, showing that AI-assisted theorem proving is no longer science fiction.
But we’re not just creating superintelligent agents to crush humans. Indeed, AlphaZero-discovered strategies improved the gameplay of four chess Grandmasters, proving that superhuman systems can teach the very best humans, not just beat them.
AI is now a lab partner too. DeepMind’s Co-Scientist and Stanford’s Virtual Lab generate, debate, and validate hypotheses, discover new and established ideas as science is becoming a closed loop with AI in it.
Biology gets its scaling laws too. Profluent’s ProGen3 trained on 1.5T tokens and created a compute frontier for protein language models. This is unlocking generalisation in novel protein space and a path to novel therapeutics such as custom gene editors.
Robots now reason too. “Chain-of-Action” planning brings structured thought to the physical world - from AI2’s Molmo-Act to Gemini Robotics. Massive amounts of effort are thrown into the mix, expect lots of progress here…
Anthropic’s Model Context Protocol is the new USB-C of AI. A single standard to connect models to tools, already embedded in ChatGPT, Gemini, Claude, and VS Code, has taken shape. But not without emerging security risks…
Now, onto the Industry section. RIP AGI, long live Superintelligence. AI pilled tech executives have rebranded the mission, and it’s working: provocative, undefined, and exciting.
The frontier fight is relentless. OpenAI still tops most leaderboards, but DeepMind’s stays there longer. Timing releases have become its own science…not least informing financing rounds like clockwork.
Capability per dollar is doubling every few months. Google’s rate: 3.4 months. OpenAI’s: 5.8 months. More predictable gains are driving more investment, and more intelligence for less money.
AI software adoption has gone mainstream. Ramp data shows 44% of US businesses now pay for AI, up from 5% in 2023. Average contract value for AI products hit $530k in 2025 and is expected to pass $1M in 2026. 12 month retention is now 80%+.
AI-first companies still outrun everyone else too, growing 1.5x faster than peers on Standard Metrics.
DeepSeek’s “$5M training run” deep freak was overblown. Since the market realised the fine print in the R1 paper, that’s led to Jevons paradox on steroids: lower cost per run → more runs → more compute needed, buy more NVIDIA.
Enter Stargate: a $500B, 10GW US mega-cluster (4M chips) backed by Altman, Masa, Ellison, and Trump. The industrial era of AI begins. What a time to be alive.
Sovereigns join the race: from China’s $5B Big Fund to the UAE’s MGX, nations are writing cheques to stay in the game. We expect some nations to just tap out and declare neutrality.
China leads in power infrastructure too, adding >400GW in 2024 vs 41GW for the US. Compute now clearly runs on geopolitics.
NVIDIA still rules research and crushes its competitors: Hopper chips surge, Jetsons rise, legacy GPUs fade. If you’d just bought NVDA stock instead of its challengers, you’d be up 12x vs. 2x.
Now, let’s switch gears into Politics. The US Government is turning capitalist. Golden shares in US Steel, stakes in Intel and MP Materials, and revenue cuts from NVIDIA’s China sales. New-age Industrial policy?
America’s new “AI Stack” exports compute, models, and compliance to allies. Open source is now national security.
The AI Safety Institute network has collapsed. Washington ditched attending meetings altogether, while the US and UK rebranded “safety” into “security.”
Europe’s AI Act is wobbling: only 3 states are compliant, leaders calling it “confusing,” and pressure mounting for a pause as it’s clear the continent is being left behind.
China’s spending through the debt as Xi told ministers to “redouble efforts” on AI, boosting science funding 10% despite record debt.
Moving into Safety: budgets are anemic. All 11 major US safety orgs will spend $133M in 2025…less than frontier labs burn in a day.
Cyber and alignment risks accelerate. Models can now fake alignment under supervision, and exploit code faster than humans fix it.
But users love AI anyway. Our new State of AI Survey of 1.2k AI practitioners shows that 95% use AI at work or home, 76% pay out of pocket, average spend keeps climbing, productivity gains are real and use cases abound.
Contribute your experience to the survey at stateof.ai/survey
Reviewing last year’s Predictions, we scored 5/10. Here is a sample of our 10 predictions for next year:
- A Chinese lab tops a global leaderboard.
- AI agents make a real scientific discovery.
- Datacenter NIMBYism hits US elections.
- Trump bans state AI laws.
You can check out the full report over on the State of AI website:
The State of AI Report is a team effort. Thanks to my collaborators Zeke Gillman, Nell Norman, and Ryan Tovcimak, and everyone who helped make this our most ambitious report yet, including our reviewers Paige Bailey, Chris Gagne, Shubho Sengupta, Philippe Schwaller, David Stutz, Divy Thakkar, Neel Nanda, Aleksa Gordic, Ross Taylor, Joe Spisak, Ido Hakimi, Ryan Julian, Xander Davies, Daniel Campos, Jacob Portes, Joyce Benaich, and Jacob Arbeid.
We hope you enjoy reading!
Please share, along with any comments, thoughts or feedback :)
very nice update. some slides on how much these ai labs and/or gen ai products are losing in the near term would help, assuming no one is turning a profit of these models yet.
The gap narrows when everyone copies the same model recipe. The next gap-maker is meta-layer integration: orchestration, tool use, memory, eval loops, and productized feedback. That’s where durable advantage lives.