Fluid: How we built serverless servers
By Steven Van ·
Vercel's engineering writeup explains the custom Lambda tunnel behind Fluid compute, now handling 45 billion requests a week.
Vercel has published a behind-the-scenes account of how it built Fluid compute, the serverless model it introduced in February. Because Vercel Functions run on AWS Lambda, which didn't support streaming responses, the team built its own secure TCP-based tunnel connecting each Lambda instance back to Vercel's infrastructure, sending HTTP responses back as a stream of packets instead of a single blob. That transport layer, built over more than two years, is what let Vercel add streaming, reduced cold starts, and its many-to-one scaling model on top of Lambda's simple request/response execution.
Fluid compute now runs alongside Active CPU pricing, introduced at Vercel Ship 2025, and together the two handle over 45 billion requests a week, with some customers saving up to 95% on compute costs and no charge for CPU time spent idle.
