Mary Meeker's 340-slide AI report: key findings distilled

Executive overview

AI companies are burning billions today to win a race toward AGI — and the economics justify it. Training costs have risen 2,400x, but inference costs have fallen 105,000x, meaning smarter models are simultaneously more expensive to build and cheaper to run.

The result: explosive user growth, surging token consumption, and a clear near-term winner in Nvidia. The long-term winner is likely the application layer.

Falling inference costs are the unlock — cheap, intelligent output drives mass adoption.

Instant expansion: OpenAI's growth vs prior platforms

• OpenAI reached the same international user base as the internet did — but in 3 years vs 23.
• Matched Google's search volume in 2 years vs 11 (5.5x faster).
• Reached 100 million users in 2 months; the next fastest took 3–4x longer.
• Adding image generation (GPT-4o, March 2025) nearly doubled users, subscribers, and revenue.
• Growth is enabled by building on existing infrastructure — the "standing on the shoulders of giants" effect.

The paradox of AI economics

• Training cost: up 2,400x since 2016 — more data, compute, and parameters required.
• Inference energy cost: down 105,000x across Nvidia chip generations — near zero per token on Blackwell.
• Inference price: down ~10x across model generations; newer, smarter models are cheaper to run than older ones.
• Infrastructure maintenance cost has dropped from ~100 to below 10 (relative index).
• Google's token processing volume grew 50x in one year: 10 trillion → 480 trillion tokens.
• Who wins now: Nvidia captures ~25% of all global data center capex spend.
• Who wins next: likely the application layer — companies building on top of models, not the models themselves.

Infrastructure build-out

• Global data center capacity 5x'd between 2020 and 2024; projected to 16x that base over the next four years.
• XAI built Colossus (100,000 GPUs) in 122 days, then added another 100,000 GPUs in 92 days.
• For comparison, building a 2,000 sq ft US home takes ~234 days.
• Energy is the primary constraint on AI progress — the country that generates the most electricity wins.
• China surpassed US electricity generation in 2010; it now adds the equivalent of the entire US grid every year.
• US electricity generation is largely static — a structural disadvantage in the AI race.

Adoption: lower than expected

• Only 37% of US adults have ever used ChatGPT; even the 18–29 cohort peaks at 55%.
• Daily usage maxes at ~20 minutes per day — far below the potential value available.
• Average session length: just over 2 minutes; max sessions per day: ~8.
• The gap between what these tools can deliver and what users actually extract is large.

Model convergence and AI company growth

• As of early 2025, top models are converging on similar benchmark scores (LMSYS Arena) — but single benchmarks miss multimodal and specialised use cases.
• Top 100 AI companies reached $5M ARR in 24 months vs 37 months for top 100 SaaS companies.
• Investor expectations will likely compress this timeline further.

Model preference and market share

• OpenAI leads consumer market share by a wide margin; Google Gemini is second.
• DeepSeek ranks above Grok, Perplexity, and Claude on website visits — driven by China and India user bases.
• Claude ranks last on consumer visits despite strong model quality — quality does not equal distribution.
• Claude's API usage (developer/code use cases) would rank it materially higher if included.

Open source vs closed source

• Open source models lag closed source by ~17 months in training compute (FLOPs).
• Open source is closing the gap in text — but multimodal models (audio, image, video) require far more compute.
• Multimodal represents a tiny fraction of current open source models on Hugging Face.
• Closed source models are likely to widen their lead as multimodality becomes the primary capability frontier.