Mary Meeker's 2025 AI Report: what to take away

Mary Meeker published her first AI report in May 2025. As with her Internet Trends reports, the document is dense, US-centered, and full of data. Here is my reading as a European practitioner.

The three charts to remember

The “Trends — Artificial Intelligence” deck (BOND, May 2025) is a steamroller of exponential curves. The thesis fits in one sentence: AI adoption is climbing faster than electricity or mobile did in their day. ChatGPT going from zero to 800 million weekly users in 17 months is not just another KPI — it is a regime change.

That said, I did not retain all 340 pages. I retained three charts.

1. Capital follows the heat. $212 billion in cloud CapEx in 2024, up 63% in a single year — mostly GPU farms. When hyperscalers commit that kind of money, it is no longer marketing: it is an infrastructure bet. Capital has already made up its mind.

2. Developers vote with their IDE. The NVIDIA ecosystem reached 6 million developers in 2025. This is the leading indicator I watch first: the tools developers choose today are the products your teams will be using tomorrow.

3. The world is mobile-first and AI-native. 90% of mobile LLM users are outside North America by year three. Growth is not coming from where we usually look.

The combined message of these three curves: gravity has already shifted. Ignore it, and your five-year plan turns into a game of roulette.

Growth is beyond doubt, margins are not

The deck is a festival of up-and-to-the-right curves. Yet Meeker herself remains cautious — and I share her reservation. Three promises of the moment deserve a reality check:

• “Inference costs next to nothing.” On the ground, power grids are running hot and cooling water is getting scarce.
• “Open source kills margins.” Compliance, support and security all carry a price — someone will end up paying it.
• “The jobs are disappearing.” The first pilots cited show +14% productivity, not social carnage.

The moral I draw: growth is beyond doubt, margins are not. Model your energy costs and your revenue cannibalization before chasing the latest chip.

Three reading grids for your 2026 budgets

1. Interface → agents → autonomy. Chat was AI 1.0; agentic orchestration is 2.0. Your products need to expose verbs, not screens.
2. Compute is not the advantage — distribution is. As inference costs collapse, the barrier shifts back to data and user trust.
3. Global by design. The next billion users arrive by voice and satellite, not the browser. Plan for real-time localization — or lose the game at “hello”.

The playbook I take from it

Concrete, not theoretical — one quarter at a time:

• Q3 2025 — run a “Model Diet” audit: sort every workflow into automate / augment / archive. The goal: free up 15% of OPEX within six months.
• Q4 2025 — stand up an AI ethics red team, with real veto power. Regulators will demand it; talent already expects it.
• Q1 2026 — negotiate long-term energy contracts, 100% renewable. Energy is the new cloud bill.
• Q2 2026 — ship a first agentic feature that generates revenue, however small. Momentum beats perfection.

And depending on where you sit in the organization:

• Entrepreneurs — speed is your only moat. Claim a niche before the giants finish their buying cycles.
• Operators — measure success in automated decisions, not dashboards viewed.
• Regulators — control compute exports, not paperwork. The genie is not going back into the lamp.
• Individual contributors — the market now pays for orchestration skills: clear prompts, agent chaining, data hygiene.

The blind spot

The report massively over-represents US vendors and under-represents the European ecosystem: open source, sovereignty, and regulatory constraints. Read it with that in mind, especially if you are making infrastructure decisions for an organization subject to the AI Act.

Move fast, question even faster

The cost of paralysis now exceeds the cost of experimentation. Treat the Meeker report as a compass, not a gospel: scrutinize every shiny slide, but feel the ground shifting under your feet too.

Your next executive committee should no longer be asking “are we doing AI?” but “how many processes did we free up or improve this quarter — and what did we learn?”. In an era where machines accelerate faster than we do, the durable advantage remains the capacity to reinvent yourself before a competitor — or an algorithm — does it for you.