Your Cloud. Your Hardware. No Virtualization Tax.
Build a private cloud on servers you own, or bring your own cloud (BYOC) and reclaim the VMs you already rent. Every workload runs in its own Linux kernel with dedicated cores: bare-metal performance, kernel-level isolation, no hypervisor, no shared kernel attack surface.
Built on upstream Linux, 100% open source
Isolation Built into the Architecture
In a traditional Linux kernel, application code and device processing compete for the same CPU cores. The split-kernel architecture removes this contention by offloading all device processing to a dedicated kernel on dedicated cores. Applications see zero device interrupts, not because of tuning, but because device interrupts physically cannot reach them.
Measured on the Same Machine
Same server, same kernel build, one variable at a time: a multikernel instance against a fully accelerated KVM guest, or two kernels against one. Every number links to its setup, caveats, and reproduction steps.
Against a guest with EPT, APICv, and pinned vCPUs. KVM closes the gap only by spinning its idle cores, at up to 19 W per guest. Multikernel stays fast while sleeping in real C-states.
lmbench vs KVM 10x Less CPU to receive at line rate 21% vs 215% of a coreBoth sides hit line rate. The VM paid 2x the CPU to send, 10x to receive, and added four times as much round-trip latency (0.40 ms vs 0.10 ms) through vhost-net.
iperf vs KVM 2.6x More throughput from the same 24 cores 780K vs 300K unlinks/sTwo kernels instead of one, on identical hardware: 2.1x on stat() of a shared file, 4.1x on unlink across sockets. Locks that are not shared are not contended. Control benchmarks moved by at most 1%.
will-it-scale 153x Lower p99 beside a noisy neighbor 232 ms vs 35.6 sA latency-sensitive task next to a memory-hungry batch job. On one kernel, cgroups protect the victim only by starving the batch job to 10 passes. With two kernels, both run at full speed.
Tail latency 61x Faster large-memory instance setup 111 ms vs 6.81 s, 32 GiB preallocatedQEMU with preallocated RAM spends seconds populating pages before a vCPU runs, and pinning memory for a passthrough device costs the same. Multikernel assigns physical ranges it already owns, in about 110 ms from 1 GiB to 32 GiB.
Instance memory setup 0 Bytes lost across four kernel panics 4 of 4 panics verified intactAn app-kernel panicked mid-write four times. Each time, a fresh kernel booted on the same DAXFS memory and found every byte in place, down to a half-written page. No log replay, no reload.
Crash recoveryBuilt for Teams That Run Their Own Infrastructure
Platform & Infrastructure Teams
Consolidate more Linux workloads per machine without a hypervisor tax, without noisy neighbors, and without renewing virtualization licenses.
AI Infrastructure Teams
Sandbox untrusted agents and generated code in their own kernels, without nested virtualization. Large-memory instances are ready in about 110 ms.
SRE & Fleet Operations
Apply kernel security patches across the fleet with zero downtime and instant rollback. A crash in one kernel stays in that kernel. No maintenance windows required.
Built on Split-Kernel Architecture
Three open-source software products built on split-kernel architecture. Download from GitHub, deploy on your infrastructure.
Multikernel Private Cloud
Private cloud operating system
Private cloud without the hypervisor. Consolidate Linux workloads on bare metal in your data center, or bring your own cloud: the VMs you already rent. Each gets its own kernel, dedicated cores, and native hardware access.
Multikernel Sandbox
AI agent sandboxing runtime
Kernel-level sandboxing for AI agents. Each agent runs in its own kernel on dedicated cores, with strong isolation. No nested virtualization required.
Multikernel LiveUpdate
Zero-downtime kernel upgrade tool
Zero-downtime kernel upgrades with instant rollback. Apply security patches and kernel updates without rebooting or interrupting running workloads.
Run Your Infrastructure on Kernels, Not Hypervisors
Multikernel is open source and runs on the hardware you already have. Start with the docs, or talk to the team that builds it.