Vercel has consolidated three previously separate infrastructure products, builds, sandboxes and serverless functions, onto a single compute layer it calls Fluid. The company disclosed the unification on its engineering blog, along with usage figures it says the system now handles: a trillion-plus requests monthly, 25 million sandboxes weekly, and north of 15 million builds every day.
The stakes are about who owns the machine-provisioning problem for AI-era workloads. Cloud vendors have spent a decade selling developers pre-shaped compute (a container here, a VM there) and asking them to fit their code to it. Vercel’s pitch inverts that: describe the work, and Fluid assembles the machine, whether that means a CPU-heavy box for compiling a build, a small fast-booting instance for a single function call, or a flexible sandbox for an agent’s own code.
That inversion matters most for agents, and this is where the announcement gets more interesting than a routine infrastructure post. An agent call does not behave like a typical web request. It fires off a tool call, then sits idle for seconds or longer while a language model generates a response, then resumes. Serverless billing was built for request-response workloads where compute and wall-clock time move together; an agent breaks that assumption, because most of its runtime is spent waiting on something else entirely. A compute layer that only charges for CPU actually consumed, rather than for the shape of that idle stretch, is the real product decision underneath this announcement, not the branding of a unified platform.
Vercel’s technical explanation for how it achieves that centers on two components. Hive, described as the underlying hardware layer, provisions isolated virtual machines that the company says are typically already warm and ready in milliseconds. On top of that, a format called VHS (Vercel Hive Snapshot) lets custom developer environments resume from a saved state instead of booting cold, a capability Vercel says its own v0 product already relies on for its development environments. A newer component, Vercel Drives, decouples storage from any single machine so a workload’s files persist across sessions; it remains in private beta and currently attaches only to sandboxes.
None of these performance or scale claims come with independent verification. The trillion-requests figure, the millisecond boot times and the build and sandbox volumes are all numbers Vercel is reporting about its own infrastructure, in a post written to promote that infrastructure. No third-party benchmark or customer-reported figure accompanies any of them, and Vercel does not name which workloads make up the bulk of that request volume.
The company also does not disclose pricing detail beyond noting that its Active CPU model bills for execution alone, so time spent waiting on a model response or a database query is not charged. That distinction is the one worth testing directly, because it is the mechanism, not the trillion-request headline, that determines whether Fluid actually costs less to run agent workloads than a conventional serverless function would.
Teams running agent pipelines on Vercel, or evaluating a move there, should benchmark actual CPU-billed time against wall-clock time on a representative agent workload before assuming the unified compute story translates into a lower bill.
Vercel detailed the Fluid compute architecture on its own engineering blog.
