The Snow Family (Compute)

At a remote site — an oil rig, a farm, a military forward base, a moving vehicle, a disaster zone — there is often no reliable network back to a Region, but there is still real compute to do: filter sensor data, run inference on a camera feed, host a local app. The AWS Snow Family (compute focus: Snowcone and Snowball Edge Compute Optimized) are portable, ruggedized, sometimes battery-powered devices that AWS ships to you. They run EC2 instances, ECS/EKS Anywhere, and Lambda functions locally, disconnected if needed, and can later carry any captured data back to AWS.

• A device is ordered from a Region and jobs are defined there, but once it ships it operates wherever you physically put it — there is no AZ, and it can run fully offline.
• Devices can be clustered (Snowball Edge) for local durability and more capacity; a cluster of 5–10 nodes tolerates node loss.
• Data and AMIs are pre-loaded from your Region before shipment; results sync back when the device returns or regains connectivity.

graph LR
ORDER["Order + build job in AWS Region (pick AMIs, data)"] --> SHIP1["AWS ships device"]
SHIP1 --> EDGE
subgraph EDGE["Remote / disconnected site"]
  DEV["Snowcone / Snowball Edge Compute Optimized"]
  DEV --> EC2["Local EC2 instances"]
  DEV --> LMB["Local Lambda"]
  DEV --> K8S["ECS / EKS Anywhere"]
  DEV -.->|"optional GPU"| ML["On-device inference"]
end
EDGE -->|"ship back / reconnect"| SYNC["Data + results sync to S3"]
classDef dev fill:#e8f0fe,stroke:#3b82f6,stroke-width:2px,color:#173a70
classDef out fill:#eaf7ec,stroke:#2e8b3d,stroke-width:2px,color:#1a3d20
class DEV dev
class SYNC out

Billing is a per-device service fee for the job, then a per-day usage charge for each day you keep it beyond the included period, plus shipping and normal S3 / data transfer for anything synced back. There’s no On-Demand EC2 rate on the device — the local instance capacity is part of the device you’re renting. Snowcone is the cheapest (tiny, ~2 vCPU, portable); Snowball Edge Compute Optimized costs more and can include a GPU.

• Compute vs storage Snow devices: Storage Optimized is for bulk data transfer / migration (covered in Migration & Transfer). Compute Optimized and Snowcone are for running workloads at the edge — that’s this topic.
• Trigger phrases: “edge location with intermittent or no connectivity,” “process data locally before sending to AWS,” “ruggedized / portable / battery-powered,” “run EC2 or ML inference at a remote site,” “harsh environment.”
• Snowcone = smallest, lightest, ~8 TB usable, can run on battery, even fits on a drone. Snowball Edge Compute Optimized = far more vCPU/RAM, optional GPU, clusterable.
• Can run EC2 AMIs, Lambda functions, and ECS/EKS Anywhere locally; supports NFS and S3-compatible endpoints on the device.
• Snow vs Outposts: Snow is temporary, disconnected-capable, and ships back; Outposts is a permanent, always-connected Region extension.
• DataSync can run on a Snowcone to move data to/from the device efficiently.