Haru, fully managed.

Haru Cloud will run Haru— Arkor's GPU orchestration layer — as a fully managed cloud. Active/Standby fleets, hot failover, and standby GPUs that train instead of idling, without operating any of it yourself.

Coming soon

Meet Haru, the open-source core

Everything below the endpoint, operated for you.

Self-hosting Haru means running its control plane: a Postgres state store, runtime supervision, and the failover machinery. Haru Cloud keeps the same architecture and takes over the operating.

Provisioning

GPU fleets come up across regions or clouds through SkyPilot placement. You describe the fleet layout; Haru Cloud sources and assembles the hardware.

Supervision

Heartbeats, VRAM verification, and synthetic probes run continuously against every model the fleet routes — not just the hosts.

Failover

The full promotion sequence — stop standby training, verify VRAM, wake vLLM, probe, flip the pointer — executes and is watched around the clock.

Same Haru. Zero ops.

Haru stays open source, and self-hosting stays a first-class path. Haru Cloud is for teams who want Haru's resilience without owning another control plane.

Self-hosted Haru

You operate it

Run the open-source control plane on your own infrastructure. Full control over the state store, the supervisor, and every placement decision. Learn about Haru.

Haru Cloud

Arkor operates it

The same fleets and failover semantics, delivered as a managed service. You bring models and traffic; provisioning, supervision, and promotion are handled for you.

Coming soon.

Haru Cloud is under active development on top of the open-source Haru core. Until it launches, you can self-host Haru today — the orchestration layer is public on GitHub.