Vercel Unveils 'Fluid,' a Unified Compute Layer Optimized for Agent Workloads

Vercel has unveiled Fluid, a single system that assembles machines tailored to workloads and swaps configurations in real time. Fluid consists of Hive (hardware), Fluid images (environments), and Vercel Drives (storage), with Functions, Sandboxes, and Builds running on the same infrastructure.

Core Components of Fluid

• Hive: A hardware layer that provisions isolated VMs in milliseconds (ms). Machines are typically pre-warmed and ready for immediate execution, with each machine isolated in a multi-tenant environment.
• Fluid images: Converts user images into an optimized boot format called VHS (Vercel Hive Snapshot). This enables resume instead of boot, preparing custom machines in milliseconds.
• Vercel Drives: Separates storage from compute to ensure data persistence. Even if the underlying compute is replaced, sessions can resume exactly from where they left off in the next session.

Workload-Specific Optimization and Performance

Fluid allocates resources based on workload characteristics. Builds are compute-bound workloads running on powerful machines with high CPU and memory, while functions are IO-bound workloads that load code immediately upon request. Sandboxes accommodate various configurations and user data for flexibility.

Currently, Fluid handles 15 million builds per day, 25 million sandboxes per week, and 1 trillion requests per month. Particularly for AI agent workloads, standard VM provisioning speeds were too slow, making even the cloud unsuitable, but Fluid solves this by providing isolated VMs instantly. Additionally, it applies Active CPU billing, so you pay only for actual code execution time, not idle time.