Dynamic Compute for Modern Workloads
Vercel has unveiled details about its new compute infrastructure architecture named Fluid, a system that currently processes over 15 million builds per day, 25 million sandboxes per week, and 1 trillion requests per month. Unlike traditional cloud hosting models that require fixed virtual machine (VM) configurations, Fluid dynamically provisions computing resources in real time to match the exact requirements of each individual workload.
Core Architecture: Hive, Fluid Images, and Drives
According to Vercel, the Fluid architecture is built upon three core components: Hive, Fluid images, and Vercel Drives. Hive serves as the proprietary hardware orchestration layer, responsible for provisioning isolated VMs in milliseconds within a multitenant environment. Rather than relying on standard containers or isolates with limited security boundaries, Hive delivers full VMs to ensure robust isolation while maintaining near-instantaneous startup times.
For execution environments, Fluid images allow developers to upload custom images to the Vercel Container Registry (VCR). The system automatically converts these into the Vercel Hive Snapshot (VHS) format, enabling VMs to resume state directly from snapshots rather than booting the operating system from scratch. This capability currently powers the v0 code generation tool for instant dev environment setup. Complementing this is Vercel Drives, a portable storage layer decoupled from the VM lifecycle, currently available in private beta for sandboxes ahead of a broader ecosystem rollout.
Tailored for AI Agents with Active CPU Pricing
Vercel noted that this architecture directly targets AI agent workloads—tasks that frequently execute arbitrary code, demand strict security boundaries, and cycle through ephemeral VM lifecycles. Rather than billing for the entire duration a VM remains connected, Fluid adopts an Active CPU pricing model, charging solely when the CPU actively processes code and eliminating costs during idle periods while waiting for database queries or AI model responses.