Are Unikernels the Answer for Next-Gen AI Cloud Workloads? (thenewstack.io)

Vercel’s venture arm led part of a $6M seed round for Unikraft, the company commercializing the Linux Foundation’s Unikraft unikernel technology; the deal also gives Unikraft access to Vercel’s CDN and cloud-running expertise. The investment follows production use by Prisma, which uses Unikraft SDK and Unikraft Cloud to run thousands of dedicated PostgreSQL instances as tiny unikernel microVMs (assembled via Docker Compose and deployed on Cloudflare Workers). Because each unikernel contains only the application and the minimal kernel, Prisma can spin database instances up in milliseconds, scale them to zero, and host many more instances per server—enabling a free tier and big cost savings compared with containerized Postgres. Technically, unikernels are single-address-space binaries (no kernel/user split) that trade generality for extreme minimalism: faster execution, smaller attack surface, lower memory usage and 50–100% higher throughput in some tests versus vanilla Linux. Unikraft also added Kubernetes integration so unikernel instances appear as nodes and can use standard k8s scaling logic. While unikernels historically struggled with tooling and POSIX compatibility, their tiny footprint, millisecond boot times and improved latency make them compelling for next‑gen, highly distributed AI/ML workloads that spawn massive numbers of short-lived services and need to eliminate cold starts and costly network hops.