The infrastructure layer is consolidating while the application layer fragments into a thousand bets on what agentic AI actually means in practice. Nvidia is spending 26 billion to build open-weight models and striking 2 billion deals with cloud providers like Nebius, essentially funding its own customer base while courts and regulators wrestle with what safety even looks like. Meanwhile, every category of software is being rewritten around agents: Zendesk bought Forethought for customer service automation, Meta acquired Moltbook to position itself for an agentic web, Databricks bought Quotient AI to help enterprises monitor agent behavior in production. The pattern is clear: whoever controls the evaluation and training infrastructure for agents controls the margin.
But the application bets are chaotic and the safety narrative is fracturing. Character.AI was deemed uniquely unsafe by researchers, yet Grammarly's AI feature was quietly shut down after a lawsuit revealed it falsely attributed suggestions to real authors without consent. Ford's new AI assistant checks seatbelt compliance in fleet vehicles while AI meal plans for teens are cutting calories and carbs in ways nutritionists wouldn't. Anthropic is launching a think tank to examine AI's societal effects while simultaneously selling Claude to the US military for targeting decisions in Iran. The gap between what these systems are actually being deployed to do and the frameworks being built around them is widening, not closing.
The money is flowing toward builders who can show revenue and production usage, not flashy demos. Lovable hit 400 million ARR with 146 employees, Replit jumped to a 9 billion valuation six months after hitting 3 billion and is chasing 1 billion ARR by year's end, and Forethought was acquired at a valuation reflecting years of head start in a category that barely existed two years ago. Yet the labor math is inverting: Atlassian cut 10 percent of its workforce citing AI threats, Oracle set aside 500 million for restructuring as it deploys AI coding tools, and Salesforce had to pay a premium on a 25 billion bond deal because Wall Street is spooked about AI disruption. The valuation premium for fast-growing AI startups with proven unit economics is real. The discount for mature software companies is real too.
Sloane Duvall