Vercel's Agent Infrastructure Bet and Stripe's AI Pricing Playbook

Vercel shipped a full stack of agent-native infrastructure (persistent storage, backends, Python, a "Token Pass" pricing primitive) in a single day, Stripe's product team published a sharp framework for why AI monetization breaks traditional SaaS pricing logic, and Simon Willison surfaced the first real number on Anthropic's business: $16.8bn in quarterly revenue, up 50%, ahead of an IPO. ### Vercel - Ships Persistent Storage, Python GA, and "Token Pass" for Agent Workloads **Source:** https://vercel.com/blog/introducing-vercel-drives-persistent-storage-for-agents, https://vercel.com/blog/python-on-vercel-is-now-generally-available, https://vercel.com/changelog/introducing-token-pass **Credibility:** High (first-party changelog and blog posts, multiple shipped features with GA status) **What happened:** Vercel dropped five related announcements on the same day: Vercel Drives (persistent storage built specifically for agents), general availability of Python on Vercel, a new "Backends" primitive supporting containers and long-running daemons, a public beta of Vercel Daemons, and "Token Pass," a new pricing/metering concept. Taken together, this is Vercel repositioning itself from "frontend deployment platform" to "infrastructure for agents that need to remember things, run in the background, and execute arbitrary code." **Key technical details:** - Vercel Drives: persistent storage designed for agent state, separate from ephemeral function storage - Python on Vercel: GA, not beta, signaling Vercel wants to be a serious runtime for AI/ML workloads, not just JS/TS - Backends: supports containers and "Daemons" (long-running background processes) alongside Python, a structural departure from Vercel's serverless-only origins - Vercel Daemons: public beta, specifically framed around agents that need to run continuously rather than request/response - Token Pass: a new pricing mechanism, details light in the changelog excerpt but named specifically around token-based usage - Also shipped same day: Bring Your Own Cloud on AWS GA, Enterprise-Managed Authorization for Connect, MCP Events support in Connect (public beta) **Why it matters for PMs:** This is Vercel admitting that serverless functions (their entire original architecture) don't fit how agents actually behave: agents need to remember state across sessions, run for hours without a request triggering them, and execute code outside JS. If you're building an agent product and evaluating where to host it, Vercel just closed the gap with AWS/GCP on the "can it run my agent" question while keeping its deploy-simplicity story. The MCP Events beta in Connect is also worth watching: it suggests Vercel wants to be the connective tissue between agents and enterprise data sources, not just a hosting layer. **Critical questions:** - What does Token Pass actually meter, and how does it compare to the token-based pricing everyone else (OpenAI, Anthropic) already uses? Is this a wrapper on top of model costs or a separate Vercel-level charge? - Does GA Python support mean Vercel is now competing with Modal, Replicate, and AWS Lambda for ML inference workloads, not just app hosting? - If agents need persistent storage and background daemons, how different is this architecturally from "just run a server," and what's left of the original serverless value prop? **Action you could take today:** If your team hosts any agent-based feature on Vercel today, check whether your current workarounds for state persistence (external DB, Redis) could be simplified by Vercel Drives, and whether Daemons removes the need for a separate background-job service.