Fluid compute: Evolving serverless for AI workloads
By Steven Van ·
Fluid reuses idle compute across requests instead of spinning up a new instance each time, and is now the default for new Vercel projects.
Fluid compute, a new model for Vercel's serverless functions, reuses existing compute capacity across requests instead of spinning up a new instance every time. Vercel built it for AI workloads specifically: an LLM call involves a function that stays running while it waits, sometimes for minutes, for a response, and traditional serverless bills that idle time and typically spawns a fresh instance for the next request even when existing ones are free.
- Fluid scales inside a single instance before creating new ones, so one instance can process multiple AI inference requests at the same time and cold starts are reduced
- Compute is placed in regions close to the underlying data to keep performance consistent
Secure by default
Requests are routed to the nearest Vercel point of presence and inspected by the Vercel Firewall before reaching the Functions router. Fluid instances aren't exposed to the internet directly; they connect to that router over secure, persistent TCP tunnels. Fluid also fails over automatically across availability zones within a region, with multi-region failover available to Enterprise customers.
Fluid compute is now the default for new projects and can be enabled for existing ones, according to the announcement.
