Mary Meeker's AI trends report: key data points on user growth, capex, and inference costs

Summary

Mary Meeker's latest AI trends report draws from over 300 slides of data on infrastructure buildout, user adoption, and cost dynamics. ChatGPT doubled from 400 million to 800 million weekly users in the past year. Developers building on Nvidia's ecosystem have surged past 6 million. These user numbers matter less for their scale than for what comes next. As internet services saturate, incremental users in developing nations have lower willingness to pay, which is why the narrative has shifted from consumer headcount to infrastructure spend and inference efficiency.

Inference costs

Inference costs have fallen in a straight line, approaching near-zero velocity. This matters because it unlocks use cases that require per-token economics to work and opens room for business models that were not viable 18 months ago.

Capex concentration

Meeker reframes the "Magnificent Seven" as the "Big Six": Apple, Nvidia, Microsoft, Alphabet, Amazon (AWS only), and Meta. She deliberately excludes Tesla, which invests roughly $11 billion annually across its entire business. Google, Microsoft, and AWS are each spending $60 to $80 billion per year on AI infrastructure. A few hyperscalers are driving the buildout. The thesis is that whoever locks in scale first locks in cost advantage and, by extension, model quality. Elon's xAI is structured off-balance-sheet partly to bypass this constraint and demonstrate independent capital deployment.

Distribution speed

LLMs achieved 90% of internet users in penetration, up from 50% at launch. Unlike the internet itself, which required physical infrastructure such as pipes, phones, and distribution networks, LLMs stack on top of existing networks. OpenAI did not need to build a device or distribution arm. They leveraged the internet as what Meeker calls the world's greatest distribution engine in history. That speed of adoption is unprecedented for technology adoption curves.

China's footprint

Deepseek surged from 0% penetration in February 2024 to roughly 10 to 15% by February 2025, then fell back. The data is ambiguous on whether this reflects pushback against Chinese models or normal churn. The real concern for investors betting on US dominance is that China's LLM footprint is rising and competing on cost and accessibility, not just model quality.

Data scarcity

Training datasets have grown at 260% annually over 15 years, but the curve is bending. GitHub contains only 10 to 100 million tokens or fewer, already exfiltrated and available offline. The largest public datasets from Hugging Face total roughly 44 terabytes, compressible to a single 50TB hard drive. Google's YouTube corpus is genuinely proprietary and hard to replicate. Microsoft's GitHub advantage is not as durable because the data is already in the wild. The frontier model race increasingly depends on synthetic data, proprietary use cases, or capital-intensive scale rather than raw internet tokens.

Model release cadence

From March 2023 (GPT-4, Microsoft Copilot, Google Bard, Anthropic Claude) to January 2025 (Deepseek R1, OpenAI o3, Alibaba Qwen 2.5 Max) spans 22 months. Model releases have become routine, each claiming marginal gains on reasoning or multimodality. ChatGPT hit 800 million weekly users while o3 remained gated and paywalled. Deepseek's free, open-weight R1 became the public's first hands-on experience with reasoning models. That asymmetry—expensive, closed reasoning from OpenAI versus free, accessible reasoning from Deepseek—shaped perception in ways that metrics alone do not capture.