Ten million virtual machines on a single server, each waking in ten milliseconds. That is the benchmark Unikraft CEO Felipe Huici demonstrated at QCon London, and it changes what we expect from cloud infrastructure. For years, the assumption has been that virtual machines are heavy, good for isolation but slow to start, wasteful when idle. Huici's team has shown that this trade-off is optional.
The demonstration rests on three technical choices that together rewrite the economics of server density. First, Unikraft uses Firecracker, the lightweight hypervisor Amazon built for Lambda and Fargate. Second, it relies on VM snapshots: a sleeping workload is frozen to disk, not kept warm. Third, the unikernel approach strips the operating system to only what the application needs. The result is a platform where a million virtual machines can sit dormant on a single commodity server and resume instantly on demand. Server density stops being about how much hardware you can pack into a rack and becomes a scheduling problem, how fast you can wake a snapshot.
For anyone managing cloud costs or building multi-tenant services, this is the practical shift. Scale-to-zero has been a promise for serverless functions, but it has rarely applied to full virtual machines with strong isolation. Unikraft's approach means you can run a dedicated VM per tenant, charge only for active compute, and still respond to a request in milliseconds. That turns the traditional container-versus-VM debate on its head. Containers win on startup speed; VMs win on security boundaries. Here, you get both.
What matters most is that this is not a simulation or a theoretical paper. Huici traced a decade of academic unikernel research to a working platform. The demo at QCon was live, on a commodity server, waking the millionth VM in ten milliseconds. That is a concrete milestone. The next question is not whether the technology works, it does, but how quickly the industry adopts a model where every tenant gets their own instantly-resumable virtual machine, and idle compute costs nothing. For organizations tired of over-provisioning or compromising on isolation, that future is already here.