One compute layer, three workload shapes

Vercel's compute platform now runs a single system underneath its three workload types: serverless functions, sandboxes, and builds. The system, called Fluid, assembles a machine tailored to each incoming workload, adjusts configuration on the fly, and absorbs traffic spikes in real time.

Fluid already processes over 15 million builds per day, 25 million sandboxes per week, and one trillion requests per month. It operates on a shared stack of three components: Hive (machine provisioning), Fluid images (runtime environments), and Vercel Drives (portable storage).

What changed in compute

The history of infrastructure is a progression from physical hardware to virtual machines: first you bought components and installed them yourself, then you rented bare metal, and then the cloud let you request any configuration on demand. For agents, even the cloud is too slow — a standard VM cannot provision fast enough to keep pace with an agent's workflow. Fluid closes that gap by making machines available near-instantly.

How Fluid matches a machine to a job

When a workload arrives, Fluid builds a machine to fit its shape. Hive provisions an isolated VM — typically one already warm — and your own image boots on top as the environment. A Drive connects and your files are immediately available, since storage was never tied to the machine in the first place.

  • Builds are compute-bound: they need a machine heavy on CPU and memory.
  • Functions are IO-bound: they load specific code and run it the moment a request lands, usually on a small VM.
  • Sandboxes are flexibility-bound: they adopt whatever configuration the task calls for, with a Drive attached for user data.

Under the execution model, many requests run on a single instance rather than each spawning its own. Work begins immediately, and with active-CPU pricing you pay only while your code executes — not while it waits on a database or a model.

The three building blocks

Hive: the hardware layer

Hive provisions the isolated machines every workload runs on. It selects the right machine for the job, maintains isolation across the multitenant environment, and operates at global scale. It gives every Vercel product a single control-plane API, so internal teams build on one foundation rather than maintaining separate infrastructure. Hive provisions a full VM in milliseconds and carries filesystem state with it, delivering stronger isolation than an isolate or container without the startup penalty.

Fluid images: the environment

Until now the operating system was fixed — you ran whatever the cloud provided. With Fluid images, you push your own image to the Vercel Container Registry and run it across sandboxes and functions. In the background, Vercel converts images into a format called VHS (Vercel Hive Snapshot), the optimized boot format behind Dockerfile deploys and custom sandbox images. Because VHS resumes rather than boots, a custom machine is ready in milliseconds. The format uses the same technology as Sandbox Snapshots, so optimizations in one product carry over to the others; v0 already uses it to build and run its own development environments.

Vercel Drives: the storage

Files live in portable, durable storage that travels with a workload instead of being stranded on a machine's disk. Storage that outlives the compute attached to it is what makes the rest of the system work: you can swap the underlying compute and pick up exactly where you left off. Drives attach to sandboxes today in private beta, and will extend to the rest of Fluid from there.

One system, not three platforms

Functions, sandboxes, and builds are different shapes of the same system rather than separate platforms. Because one compute layer underpins every product, any improvement in boot time, isolation, scheduling, or caching applies across all three at once — new compute shapes can be added without rebuilding the foundation, and the primitives never drift apart.

Why agents need this

An agent runs untrusted code, so it needs a strong security boundary. It brings its own tools, so it needs a custom environment. And because it constantly spins up and tears down, its state must outlive any individual machine. Fluid assembles the right machine immediately, and then a fresh one for the next task.

For years the machine was fixed, and you shaped your work around it. In Fluid, you describe the work and the machine forms around it.