Another example of how AI disrupts everything, from UIs to infrastructure: serverless. You can’t take tomorrow for granted… The model that Lambda pioneered was 1 invocation 1 function instance. This works wonderfully if your databases and caches respond instantly, as they should! This was the norm for SaaS and e-commerce. When GPT APIs came out and developers saw the opportunity to create AI products, two new profiles emerged: ▪️ Streaming responses ▪️ Lots of idle time & long responses Luckily for us, we had been working hard on streaming in preparation for React Server Components. We knew that for complex frontends with lots of data and components, the best model was to render and return as soon as possible. That investment in streaming paid off. Idle time posed a significant challenge. A system optimized for fast computation suddenly had workloads that were taking 30s to respond. And then reasoning models come out, and we’re now talking minutes. And then agents come out, and we’re talking hours? Days!? Our team moved extraordinarily quickly while operating at very large scale. We created 🌊 Fluid which solves the idle time problem by sending more requests to existing function instances. Like a VPS would. But unlike a VPS, if the instance is overloaded and blocked by compute, it intelligently launches a new instance in parallel to accommodate. Think of this as just-in-time, invocation-driven, dynamic load balancing. The customer win concretely is that we’ve saved people a lot of money. Up to 90% cost reductions for some workloads by just turning Fluid on, which, unlike other solutions, requires no code changes and runs full, off-the-shelf, versioned Node.js and Python. We’re very privileged to host many of the most popular AI apps in the world. With one of them we got to see Jevons Paradox up close while in beta. We saved them a lot of money, we scaled with them as the service took off, they continued to innovate and launch new things on Vercel. The savings over time just translated into more service demand, because we helped them win. We’ve got a few more things coming to Fluid that I think you’ll love, especially for building these kinds of AI apps. You can read more about this topic here: https://t.co/wzzQfSwaBY