Vercel Ship 2025: The Building Blocks of an AI-Native Platform

Vercel’s third annual Ship conference in New York City laid out the company’s roadmap for shifting from a frontend cloud to an "AI Cloud." The announcements centered on a core principle: as AI reshapes application development, the compute, security, and deployment models need to evolve to support agents and high-performance workloads without compromising on speed or safety.

The keynote highlighted a series of platform enhancements designed to give developers greater control over model access, background execution, and incremental releases, all while making infrastructure costs more granular. Here’s a breakdown of the major technical updates and how they fit into the broader Vercel architecture.

A Unified Endpoint for Multi-Provider AI

The AI SDK previously simplified AI integration by abstracting vendor-specific APIs. Now, AI Gateway extends that abstraction to the network level. It provides a single endpoint for accessing a wide range of models across providers like OpenAI, Anthropic, Google, and xAI, with usage-based billing at list prices or Bring-Your-Own-Key support.

This allows developers to switch providers with a single line of code, set up intelligent routing, and configure fallbacks to improve uptime. It also brings improved observability, including per-model usage and latency metrics, eliminating the need for infrastructure rewrites when newer models are released.

Rethinking Serverless Pricing with Active CPU

Vercel argues that traditional serverless platforms are poorly suited for I/O-bound tasks like AI inference and agent operations, which scale instantly but often sit idle between network calls. Their Fluid compute system addresses this by orchestrating compute across invocations, allowing concurrent requests to share resources—yielding up to 85% cost savings for some teams.

Building on this, the new Active CPU pricing model separates costs into three distinct buckets:

  • Active CPU: Charges only for the time your code is actually executing.
  • Provisioned Memory: Covers waiting time, billed at 1/11th the rate of Active CPU.
  • Invocations: A per-function-call metric standard to serverless platforms.

For example, an AI call that takes 30 seconds to respond but only uses 300ms of compute would bill for those 300ms of CPU time plus the discounted memory rate for the remainder. This model coincides with expanded defaults for Fluid compute functions, which now have a default execution time of 300 seconds, up from the previous 60–90 seconds, with higher memory and CPU allowances.

Running Untrusted Code with Vercel Sandbox

To support the rise of AI agents that generate executable code, Vercel introduced Sandbox. It provides isolated, ephemeral microVM environments for running untrusted code securely. Designed for Node.js and Python, the environments scale to hundreds of concurrent instances with execution times up to 45 minutes.

As a standalone SDK, Sandbox works outside the Vercel platform and allows developers to stream logs, install dependencies, and control runtime behavior. It was developed internally for the v0 tool and is now publicly available, priced under the Fluid Active CPU model.

Safe Deployment: From Rolling Releases to Microfrontends

Deploying rapidly requires safety nets. Rolling Releases are now generally available, offering automated, incremental rollouts to user subsets without custom routing. This feature includes built-in monitoring for metrics like Time to First Byte (TTFB) and error rates, with the ability to pause or abort the rollout if issues arise.

Release stages are configurable per project, with global updates propagating in under 300ms. Management is flexible, accessible via the REST API, CLI, dashboard, or Terraform.

For larger organizations, Microfrontends is entering limited beta. This feature allows teams to split large applications into independently deployable units, each using its own tech stack. Vercel handles integration and routing, allowing teams to modernize legacy systems incrementally while maintaining a unified view in the Vercel dashboard.

Security: Invisible Bot Detection and Agentic Insights

As bot sophistication increases, traditional header checks and rate limits fall short. BotID, built in partnership with Kasada, acts as an invisible CAPTCHA for critical routes like checkouts, logins, and expensive API calls. It injects lightweight, obfuscated code that evolves on each load to resist replay and tampering, performing deep signal analysis without user friction.

Setup requires no configuration files: install the package, set up rewrites, and mount the client. Traffic appears in the Firewall dashboard, filterable by verdict, user agent, country, or IP address.

Complementing these security controls, Vercel Agent is a new AI assistant (in limited beta) built into the dashboard. It analyzes performance and security data to summarize anomalies and recommend specific actions—such as adjusting firewall rules during traffic spikes or identifying optimization opportunities.

Background Work and Long-Running Tasks

Vercel Queues is now in limited beta, allowing developers to offload slow or long-running jobs to the background. This ensures tasks like AI video processing or external service updates can complete reliably without blocking user requests. The system uses an append-only log for message storage, guaranteeing that tasks are persisted and never lost before completion.

Together, these updates are designed to create a platform where infrastructure manages itself: handling cost-efficient inference, isolating agent-generated workloads, and shifting traffic to stable versions automatically. It signals a move toward a platform built for developers and AI agents alike.