AI Trends

Mary Meeker published a 340 page AI Trend report for 2025, May 2025.

Rapid Acceleration and Historical Parallels: AI adoption is faster than the internet/mobile eras across metrics (e.g., ChatGPT reached 100M users in ~2 months vs. Instagram’s 2.5 years). Meeker frames AI as a “meta-technology” enabling broader innovations, akin to the printing press or internet, but compressing timelines from centuries to years.

Global GDP growth charts suggest AI could drive the next economic leg up but it was tough for the internet to prove causation of the productivity gains.

Developer and Ecosystem Growth: Explosive increases in developers to 6 million.

AI startups (>27K), tools (>70K), and GitHub repos highlight proliferation. Patents in computing/AI are booming, signaling a new innovation wave.

Performance and Capabilities

AI models are improving exponentially (150% annual growth in supercomputer performance; 200% in compute from algorithms).

AI Compute has scaled another hundred times from 2023.

Benchmarks show surpassing human baselines (e.g., Turing test passed; realistic images/audio via tools like 11 Labs). Current uses: PDFs/code generation/interview prep.

By 2030 (per ChatGPT query): Human-level text, full-length films, natural speech.

By 2035: Scientific research, immersive worlds. Transcripts highlight gaps (e.g., adaptive learning, memory, intent over time) and breakthroughs needed.

2035 predictions (e.g., protein folding, drug discovery) are already emerging (e.g., AlphaFold). Risks include hallucinations, but benefits like cancer detection (faster/more accurate) and R&D timeline reductions (30-80%) are emphasized.

Usage, Adoption, and Engagement: AI chatbots (proxy: ChatGPT) see rising daily/weekly use across ages/groups, carving ~20 minutes/day from users (inflecting up with better models). Retention improving toward Google Search levels.

At work: Boosts quality/speed (e.g., production output, sales).

At school: Widespread (e.g., on laptops), raising critical thinking concerns.

Expansion to “deep research” (automating specialized knowledge) and agents (e.g., multi-step workflows), though early implementations are “sketchy.”

Enterprise adoption: Prioritizing revenue (e.g., sales/customer success) over costs (surprising, as anecdotes suggest cost-focus); case studies include Bank of America (virtual assistants), JP Morgan (end-to-end AI), Kaiser (note-taking), Yum Brands (kitchen optimization).

Capex, Compute, and Infrastructure: Big Tech capex surging (e.g., $100B+ model training; data centers as “electricity guzzlers”). Yet efficiency gains are massive (e.g., 50,000x per-unit energy since 2016;

Inference costs down 99.7% in 2 years vs. light bulb’s 75 years). Training data/compute scaling (2400x growth), but questions on limits (e.g., data exhaustion post-10^13 words).

Multi-cloud world (AWS, Azure, Oracle, Alibaba);

AI catalyzes cloud migration.

Monetization and Economics: Revenue scaling (e.g., ChatGPT subscribers/revenue vertical), but not as fast as losses/capex (disconnect to close). Pricing shifts: $200/month acceptable (e.g., Gemini, Claude); OpenAI testing $2,000.

Deepseek shows a move from upfront capex (training) to opex-like inference (profitable via tools like deep research). Deepseek uses the same training compute for inference compute to generate profits.

10x price-to-sales for AI firms, but revenue follows capex.

Physical World and Robotics: AI in robotics/autonomy accelerating.

There is reason for AI optimism. There is no plateau. AI solves real problems.

Pessimists are concerned about extinction, job loss and imminent plateau claims. Pessemists sound smart and cautious but optimists make money.