Our container platform is in production. It has GPUs. Here’s an early look

This post introduces Cloudflare's new container platform, built internally to run resource-intensive workloads like AI inference and browser rendering. It details the platform's global scheduling capabilities, which leverage Workers, Durable Objects, and KV to dynamically place containers across Cloudflare's network based on capacity and hardware availability, including GPUs. The platform is runtime-agnostic, supporting gVisor, Firecracker, and QEMU, with ongoing evaluation of cloud-hypervisor. It addresses challenges in scheduling large AI models with specific GPU memory needs and discusses optimizations for image pulling and startup times.