2025年人工智能现状报告（英）

State of AISameer DholakiaPARTNERMike DroeschPARTNERLauri MoorePARTNERKent BennettPARTNERTalia GoldbergPARTNERJanelle TengPARTNERCaty ReaINVESTORLindsey LiINVESTORMaha MalikINVESTORBhavik NagdaINVESTORLance Co Ting KehVENTURE PARTNERDarsh PatelINVESTORSource: Bessemer research from portfolio data $0M$10M$20M$30M$40M$50M$60M$70M$80M$90M$100M123456789$ ARR in MillionsYears to $100M The new $100M ARR growth curve for AI startupsAI SUPERNOVACLOUD CENTAURAI SHOOTING STARSUPERNOVASSHOOTING STARSCLOUD CENTAURAVERAGE RATES OF GROWTHYear 1 - $40M ARRYear 2 - $125M ARRYear 1 - $3M ARRYear 2 - $12M ARRYear 3 - $40M ARRYear 4 - $103M ARROn average, the top private cloud companies reach $100M ARR in < 7 years. 1.5 years to $100M$125M ARR on average by year 2$40M ARR on average in 1 year$40M ARR on average in 3 years$103M ARR on average in 4 years4 years to $100M7 years to $100MAI NOTEAI benchmarks: Supernovas vs. Shooting Stars AI BENCHMARKSWhat great AI startups look like in 2025 SUPERNOVASSHOOTING STARSYear 1 $40M ARRYear 2 $125M ARRGross margins~25%*ARR/FTE$1.133MYear 1$3M ARRYear 2$12M ARR Year 3$40M ARR Year 4$103M ARR Gross margins60%ARR/FTE$164K*For AI Supernovas, gross margins are often negative PredictionsThe browser will emerge as a dominant interface for agentic AIPREDICTION 1As agentic AI matures, the browser is emerging as the dominant interface—not just for navigation, but for end-to-end autonomous execution.Unlike voice, browsers offer a contextual, ambient surface embedded in daily workflows. Tools like Comet and Dia preview a future where AI runs natively at the operating layer—automating tasks, reasoning across tabs, and executing high-level user intent. We expect new AI-native browsers from OpenAI, Google, and others to push this frontier. The next browser wars are already underway.2026 will be the year of generative videoPREDICTION 22024 was the breakout year for generative images, 2025 for voice—and 2026 is shaping up to be video’s moment. Advances in model quality (Veo 3, Sora, Kling, Marey) and falling costs are collapsing video’s complexity barrier, driving mainstream adoption. We also expect the next 12 months to clarify the market structure for generative video: Do large labs win it all? Will open-source catchup? Are there advantages to real-time and low latency use cases? From cinematic storytelling to customer engagement, generative video is poised to reshape entertainment, marketing, education, and beyond.Evals and data lineage will become a critical catalyst for AI product developmentPREDICTION 3Most companies still can’t assess model performance in their real-world contexts. Public benchmarks fall short—missing nuance around workflows, compliance, and business-critical use. Enterprises don’t just want performance—they want confidence. That requires reproducible evaluation tailored to their own data and risk profile. As model performance converges, the real edge is knowing when, where, and why your model works. That’s why in 20