Skip to content

GPU Providers

A GPU provider is the cloud Socaity rents physical accelerators from to run your job. Socaity is the orchestrator that schedules the work, bills you for active time, and forwards container images to whichever provider you select. The provider is the layer that owns the GPU silicon, the data center, and the residency boundary; Socaity is the layer that owns the API contract and the job lifecycle.

The PROVIDER enum has six values: auto, localhost, socaity, runpod, scaleway, and azure. Working today: localhost and runpod. Reserved for future releases: scaleway and azure; both raise NotImplementedError at backend resolution. The socaity value routes a job through the Socaity orchestrator itself.

Why providers exist

No single cloud owns every GPU SKU, every region, and every certification profile you might need. RunPod has the broadest NVIDIA SKU coverage and the lowest per-second prices. Scaleway is an EU-headquartered cloud with ISO 27001 certified data centers in France. Azure offers H100 capacity and enterprise procurement. Socaity puts them behind one job API as interchangeable execution surfaces, so you can match a workload to the right surface without rewriting the integration.

The same image runs on any wired-up provider. Socaity hands the container to the provider, the provider runs it on a GPU, and Socaity returns the result through the same Job handle you already use. Switching providers is a flag change at deploy time, not a rewrite.

Provider catalog

RunPod
Working today

Serverless and dedicated NVIDIA GPUs across a broad SKU range. The current execution surface for every cloud Socaity job.

Serverless GPU
Wide SKU range
Per-second billing
Scaleway
Planned

EU-headquartered cloud with ISO 27001 data centers in France. Reserved as a provider enum value; the backend resolver raises NotImplementedError today.

EU-headquartered
ISO 27001 footprint
Enum reserved
Azure
Planned

Microsoft Azure GPU VMs as a future execution surface. Reserved as a provider enum value; the backend resolver raises NotImplementedError today.

Enterprise procurement
EU regions
Enum reserved

Tradeoffs at a glance

Status reflects the APIPod backend resolver today. Capability rows describe what each provider is positioned for; concrete numbers (per-second prices, GPU SKU availability, region coverage) live on each provider's own pricing page because they change faster than docs can. For Socaity-routed rates, see socaity.ai/Pricing.

DimensionRunPodScalewayAzure
Status in APIPodWorkingPlanned (NotImplementedError today)Planned (NotImplementedError today)
Best fitServerless GPU jobs, scale-to-zero workloads, broad SKU coverageEU-headquartered compute, GDPR-aligned procurementEnterprise procurement and existing Azure commitments
Compute modesServerless and dedicatedPlannedPlanned
GPU catalogPicked in the RunPod dashboard at endpoint creationPlannedPlanned
Per-second pricing referencesocaity.ai/Pricing for Socaity-routed rates; RunPod console for direct ratesPlannedPlanned
Region selectionPlatform-selected (EU); --region flag accepted but not yet wiredPlannedPlanned
Where it lives in the enumPROVIDER.runpodPROVIDER.scalewayPROVIDER.azure

GPU selection

Socaity does not maintain an in-product GPU SKU catalog. Each provider exposes its own hardware fleet, and you pick a SKU at the provider's surface (the RunPod dashboard, the Scaleway console, or the Azure portal) when you create the endpoint that consumes a Socaity-built image. The Socaity job model is hardware-agnostic: a model with sufficient VRAM runs the same way regardless of which NVIDIA generation backs it.

Hardware fit is bounded by model VRAM, not by API quotas. A 7B-parameter language model fits on a 16 GB card; a 70B model needs an A100 80 GB or H100. Pick the smallest GPU that holds your weights with headroom for the batch you intend to run. The socaity.ai/Pricing page lists currently-routed SKUs and per-second rates.

Regions and data residency

Region is a provider attribute, not a Socaity attribute. Today, region is selected by the platform (RunPod EU); the --region flag on apipod is accepted but not yet wired to provider routing, so per-workload selection is on the roadmap. APIPod itself does not enforce a region whitelist. Data residency follows the provider's own data-center footprint and certifications.

ProviderWhere regions are chosenData-residency footprintStatus
runpodRunPod EU, selected by the platform; --region flag not yet wired.
Per region (US, EU, APAC)
Working
localhostWherever your host machine runs. Region is ignored.
Host-defined
Working
scalewayPlanned. Scaleway operates EU data centers (France, Netherlands).
EEA (when wired)
Planned
azurePlanned. Azure operates global regions including EU.
Per region (when wired)
Planned

Which provider to pick

You want a serverless GPU endpoint today

Use

runpod
. the only cloud provider wired up today. Pair with --compute serverless to scale to zero between requests.

You are iterating locally before paying for GPU minutes

Use

localhost
. the hard default. The FastAPI router binds to 0.0.0.0:8000 and runs the same image you will deploy.

You want APIPod to choose based on --compute

Use

auto
. resolves to RunPod under serverless and to FastAPI under dedicated. One image, two deploy modes.

You need ISO 27001 EU-headquartered hosting

Use

scaleway
. planned. The provider enum value exists, but the APIPod resolver raises NotImplementedError today. Stay on RunPod EU regions while we wire it.

You need Azure procurement or enterprise commitments

Use

azure
. planned. The provider enum value exists, but the APIPod resolver raises NotImplementedError today. Talk to us if Azure is a hard procurement requirement.

Selecting a provider

The provider is selected at deploy time. Three places set the value, in this precedence order: an explicit --provider flag on apipod, the APIPOD_PROVIDER env var, and the hard default localhost. Authentication keys are separate: SOCAITY_API_KEY authenticates the orchestrator for SDK and Socaity-routed calls; RUNPOD_API_KEY is read by the SDK when you bypass the Socaity backend with a direct RunPod service address.

terminal
# Working today
socaity build --provider runpod --compute serverless
socaity start --provider localhost

# Reserved enum values (raise NotImplementedError today)
# socaity build --provider scaleway
# socaity build --provider azure

Next steps