Modal vs Spheron

Run Python functions on NVIDIA GPUs without provisioning instances; cold starts in seconds

Pay for actual GPU runtime at sub-minute granularity, with scale-to-zero by default

Define environments in code, with automatic image building and caching

From T4 and L4 through A100, L40S, H100, H200 and B200

Access GPU capacity from multiple cloud providers and certified data centers from one account, without separate signups per provider

Instances are billed by the minute rather than rounded up to the hour, past a 20-minute minimum runtime (60 minutes for a 2x AMD MI300X)

NVIDIA instances are typically ready in under two minutes with no approval workflow or provisioning queue

Compute spend across every connected provider is tracked in a single dashboard

Each VM or bare metal instance includes NVMe SSD storage, network bandwidth and full root access, with a dedicated IP on NVIDIA instances

Spheron sources large clusters, specific hardware and InfiniBand configurations from its partner data center network, with a typical quote turnaround of 24-48 hours