apiCommonsFinops: '0.1'
provider:
  name: NVIDIA NIM
  id: nvidia-nim
  url: https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
  sources:
  - https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
  - https://docs.nvidia.com/nim/index.html
  - https://www.nvidia.com/en-us/data-center/dgx-cloud/
billingSurfaces:
- id: nvidia-ai-enterprise-subscription
  name: NVIDIA AI Enterprise Subscription
  type: subscription
  description: Per-GPU per-year licensing for self-hosted NIM in production. The unit is the physical NVIDIA GPU
    on which a NIM container runs. Cost recovery is via NVIDIA invoice and is independent of inference volume.
  unit: gpu-year
  currency: USD
- id: dgx-cloud
  name: DGX Cloud
  type: usage
  description: Managed multi-node GPU service. Billed by instance-hours through NVIDIA or the hosting hyperscaler
    marketplace; NIM hosted endpoints are bundled.
  unit: instance-hour
  currency: USD
- id: build-nvidia-com-credits
  name: build.nvidia.com Credits
  type: credits
  description: Free developer credits on the hosted endpoint at integrate.api.nvidia.com. Granted on signup and
    optionally topped up for paid prototyping tiers.
  unit: request
  currency: credit
- id: cloud-marketplace-passthrough
  name: Cloud Marketplace Pass-through
  type: marketplace
  description: When NIM is consumed through AWS, Azure, GCP, or OCI marketplaces (GKE, EKS, AKS), charges flow
    through the hyperscaler bill alongside the underlying GPU instance cost.
  unit: gpu-hour
  currency: USD
focusMapping:
  AccountName: NVIDIA Enterprise Account
  ChargeCategory: Usage | Subscription
  ChargePeriodStart: invoice_period_start
  ChargePeriodEnd: invoice_period_end
  CommitmentDiscountType: annual | perpetual | committed-use
  PricingUnit:
  - gpu-year
  - gpu-hour
  - instance-hour
  - request
  ServiceCategory: AI and Machine Learning
  ServiceName: NVIDIA AI Enterprise | NVIDIA NIM | DGX Cloud
  SubAccountName: workspace | gpu-cluster | project
  SkuId: nvaie-* | dgx-cloud-* | nim-*
  EffectiveCost: invoice_line_amount
  BilledCost: invoice_line_amount
  ListCost: catalog_price
notes:
- Public LLM inference pricing is NOT published per-token by NVIDIA on the hosted endpoint — the hosted endpoint
  is a developer convenience and is metered in credits, not USD.
- Production cost is dominated by the underlying GPU bill (whether NVIDIA AI Enterprise license + on-prem GPUs,
  cloud GPU instances + license, or DGX Cloud).
- Operators model per-token cost themselves by dividing (GPU $/hour) by measured tokens/sec/GPU for the chosen
  TensorRT-LLM / vLLM engine and model.
- /v1/metrics provides the per-engine throughput and queue-depth signals required for FinOps unit-cost
  attribution.
modified: '2026-05-25'