Compute Platform

KVM, and why it matters

Our hypervisor is KVM — the virtualisation layer built into the Linux kernel, the same technology most large clouds run underneath their own branding. Running it without a proprietary wrapper has a practical consequence: a machine image that runs here runs on any KVM host, so moving in or out is a copy operation rather than a conversion project.

The hosts

Compute hosts are multi-core with terabytes of RAM and multiple network paths into redundant switching, so no single link failure takes an instance offline. We run Ubuntu LTS across hosts, and UEFI with TPM support means modern Windows Server guests install and run properly rather than in a legacy compatibility mode. Mixed estates do not need a second provider.

Virtual machines, not shared containers

An instance is a real virtual machine with its own kernel, not a container sharing one. That matters for workloads with kernel-level requirements, for security isolation between tenants, and for anything you intend to lift from physical hardware without rewriting it first.

Storage and network attached to compute

Instances attach to block volumes from a per-zone Ceph/PetaSAN cluster with triple replication — described under storage architecture — with the volume local to the zone the instance runs in, so the read path is short. Networking is private by default: your own VPC, reached over WireGuard rather than an exposed public port, described under networking and VPN. The fabric underneath is at network fabric.

Launching and protecting instances

Prepared images and custom golden images are covered under templates and images, and every plan includes snapshots and backups so a bad migration is a rollback rather than a rebuild. The NOC watches host health continuously — see status and monitoring.

When you need something else

Accelerated workloads run on dedicated cards with PCIe passthrough — see GPUs for AI/ML. Hardware you already own can sit alongside rented capacity in one environment via colocation and hybrid cloud. Plans and inclusions are on pricing; a first deployment is walked through under getting started.

Bare metal servers in Atlanta, Georgia

Some workloads should not share a hypervisor: latency-sensitive databases, licence-bound software, and GPU training that needs the whole device. For those, bare metal servers in Zone1 — Alpharetta, GA, in the Atlanta metro — give you the full machine, in a facility MarQi Cloud operates.

Dedicated vCPU cloud: what “dedicated” has to mean

A dedicated vCPU is only dedicated if nobody else is scheduled onto it. MarQi Cloud allocates vCPU and RAM from a pool you reserve at $2 per vCPU and $2 per GB of RAM per month — you scale that pool up or down from the client portal, and what you reserve is what runs your workload.