Concepts

Platform Primitives

Instance Types

Spheron GPU offerings differ on two dimensions: interruptibility and hardware isolation.

Spot instances run as VMs and can be reclaimed by the provider at any time. Dedicated instances carry a 99.95% SLA and are not reclaimed once deployed.

Hardware isolation applies only to Dedicated instances, which come in two sub-types:

• VM: Isolated virtual machine on shared physical hardware. The default for most GPU offers across all providers.
• Bare Metal: Full physical server with no hypervisor. GPU count varies by offer and provider, from single-GPU up to multi-GPU servers. Identified by the BAREMETAL suffix in the GPU type name on the dashboard.

Multi-GPU Bare Metal offers with an SXM form factor use NVLink or NVSwitch between GPUs, which is what distributed training depends on.

All instance types enforce a minimum runtime before manual termination, 20 minutes by default. Some machine types carry a longer provider minimum, shown in the wizard before you deploy.

CPU Nodes

Not every job needs a GPU. CPU Node is a CPU-only instance for build steps, data preparation, schedulers, API workers, and control planes, priced from $0.09 per hour.

CPU Node has its own page in the dashboard sidebar under Compute, with its own wizard. Every provider's CPU offer collapses onto a single option, so there is no catalog to browse. Sizes run from 2 vCPU with 8 GB of memory up to 52 vCPU with 160 GB across Verda, Spheron AI, and Massed Compute, with spot pricing available on Verda.

Through the API, CPU Node is a single offer row with gpuType: "CPU" and gpuCount: 0. Filter for it with ?computeType=cpu.

Containers

Spheron Con runs a public Docker image on a GPU or a CPU instead of a full virtual machine. Start from a template such as PyTorch, vLLM, Ollama or Jupyter, or any public image, and the container comes up with SSH and public ports, usually within a few minutes.

In the dashboard, containers are the Dedicated Containers instance type in Deploy GPUs and Deploy CPU, next to Spot VMs and Dedicated VMs. GPU containers can be stopped to pay only for /workspace storage. Through the API, a container offer's gpuType ends in _CONTAINER; filter for containers with ?computeType=container.

Instance Lifecycle

An instance does not have to be destroyed to stop costing you money. Stopping releases the GPU while keeping the disk and its contents, and billing drops to a stopped rate that covers the retained disk. Starting brings it back on the same disk with the same public IP. Restarting reboots in place without changing what you are billed.

Support varies: stopping works on Spheron AI, Spheron ES, and Spheron Con GPU containers, restarting works on Spheron AI (depending on region), Spheron MS, Spheron AM, Spheron Con, and Massed Compute, and the two lists do not match. Reserved and bare metal instances support neither.

Through the API, a stopped instance reports status: "stopped" and is part of the live set.

Regions & Providers

Spheron sources capacity from seven providers: Verda, Spheron AI, Spheron ES, Spheron MS, Spheron AM, Spheron Con (containers), and Massed Compute. GPU availability, CPU Node availability, volume support, lifecycle controls, and startup scripts vary by provider.

Most GPUs are NVIDIA. Spheron AM supplies AMD Instinct MI300X with 192 GB of HBM3e per GPU, which runs ROCm rather than CUDA. Intel Gaudi2 is available through Spheron MS.

CPU Node runs on Verda (Finland 1, Finland 3), Spheron AI (Norway 1, Canada 1, Noida, Chennai), and Massed Compute (US regions).

Persistent volumes are supported on:

• Verda (EU Finland: Finland 1, Finland 2, Finland 3): up to 10 TB, NVMe_Shared NFS-backed, can attach to multiple instances within the same datacenter region
• Spheron AI (Canada 1, Montreal 1, Norway 1, Oslo 1, US, Noida, Chennai): block volumes up to 50 TB, up to 10 per instance; in Noida and Chennai, shared volumes up to 2,048 GB, one per instance, chosen at launch
• Spheron ES (EU North 1, EU West 1, ME West 1, US Central 1): up to 256 TiB virtiofs shared volumes, multi-instance attach
• Spheron MS (US, Canada, Norway, Poland, Israel, India, Japan, Singapore, Australia): up to 16 TB block volumes, one per instance, bound at instance launch

Spheron AM and Massed Compute do not support volumes.

Networking

Each instance gets a dedicated public IP for its deployment lifetime. There is no shared NAT or port forwarding. The IP address and SSH connection command are shown in the instance details panel in the dashboard.

All ports are open by default. Access unauthenticated services (Jupyter, TensorBoard) via SSH tunneling or VS Code Remote port forwarding. Do not expose unauthenticated services directly on the public IP.

Teams

Teams let multiple users share a credit pool, SSH keys, and volumes. Three roles are available: Owner, Admin, and Member. Only Owner and Admin can manage billing. For programmatic access with team scoping, see the API Reference.