specification: FinOps Framework
specificationVersion: '1.0'
schema: https://www.finops.org/framework/
alignedWith:
  framework: FinOps Foundation Framework
  frameworkUrl: https://www.finops.org/framework/
  dataSpec: FOCUS
  dataSpecVersion: '1.3'
  dataSpecUrl: https://focus.finops.org/focus-specification/v1-3/
provider: Nvidia
providerId: nvidia
publisherName: NVIDIA Corporation
serviceCategory: AI Infrastructure
created: '2026-05-04'
# Provenance stamped 2026-08-11: this artifact was written by the API Evangelist
# bulk sweep dated 2026-05-04, not harvested from the provider. See roadmap#35.
method: generated
modified: '2026-05-05'
reconciled: true
tags:
  - GPU
  - AI
  - Machine Learning
  - Computing
  - Graphics
  - FinOps
  - Cost Management
  - FOCUS
description: 'FOCUS-aligned FinOps for NVIDIA: hybrid commercial model spanning per-GPU/year AI Enterprise
  subscriptions (sold direct or via channel), hourly cloud marketplace SKUs (AWS/Azure/GCP), DGX Cloud
  hosted GPU rentals, and free build.nvidia.com trial credits. Cost allocation is GPU-centric: licenses,
  cloud-instance hours, and inference token throughput.'
sources:
  - https://www.nvidia.com/en-us/data-center/products/ai-enterprise/
  - https://build.nvidia.com/
  - https://catalog.ngc.nvidia.com/
principles:
  - name: Visibility
    description: For AI Enterprise direct contracts, track per-GPU license entitlement vs. utilization.
      For cloud marketplace consumption, pull FOCUS-aligned cost exports from AWS/Azure/GCP. For build.nvidia.com,
      track per-API-key request and credit consumption via the NVIDIA developer dashboard.
  - name: Allocation
    description: Tag inference workloads (Kubernetes namespaces, model name, team) so per-GPU AI Enterprise
      cost can be split across consuming products. For build.nvidia.com, allocate per-API-key.
  - name: Optimization
    description: Self-host NIM under AI Enterprise once steady-state inference exceeds the break-even
      vs hosted endpoints. Right-size GPU type (L4/L40S/H100) per model size. Use TensorRT/Triton dynamic
      batching to lift GPU utilization. For training spikes, burst to DGX Cloud or cloud marketplace rather
      than buying additional perpetual capacity.
  - name: Accountability
    description: Designate per-cluster GPU owners for utilization targets (>= 60% on production AI Enterprise
      GPUs is a common SLO). Review monthly cloud-marketplace burn against forecast; flag idle GPU instances
      within 24 hours.
billingModel:
  pricingCategory: Subscription (Per-GPU) + Pay-As-You-Go (Cloud Marketplace) + Free Trial
  billingFrequency: Annual (subscription) or Hourly (cloud)
  billingCurrency: USD
  chargeCategories:
    - Purchase
    - Usage
    - Tax
    - Credit
    - Adjustment
  chargeFrequency: Recurring
focusColumns:
  ServiceName: NVIDIA AI / Developer Platform
  ServiceCategory: AI Infrastructure
  ProviderName: NVIDIA
  PublisherName: NVIDIA Corporation
  InvoiceIssuerName: NVIDIA Corporation (or AWS/Azure/GCP for marketplace)
  PricingCategory: Subscription
  PricingUnit: gpu
  BillingCurrency: USD
  ChargeCategory: Purchase
meters:
  - name: ai_enterprise_gpu_subscription
    description: Per-GPU per-year AI Enterprise software subscription
    unit: gpu-year
    aggregation: sum
    dimensions:
      - gpu_model
      - cluster
      - region
  - name: ai_enterprise_marketplace_hours
    description: AI Enterprise hourly metered consumption on cloud marketplaces
    unit: gpu-hour
    aggregation: sum
    dimensions:
      - cloud
      - region
      - gpu_model
  - name: dgx_cloud_instance_hours
    description: DGX Cloud / DGX Cloud Lepton hosted GPU instance hours
    unit: instance-hour
    aggregation: sum
    dimensions:
      - cloud
      - region
      - instance_type
  - name: build_api_credits
    description: build.nvidia.com NIM endpoint credit consumption
    unit: credit
    aggregation: sum
    dimensions:
      - model
      - api_key
  - name: inference_tokens
    description: Inference tokens generated (input/output) on hosted or self-hosted NIM (operational telemetry)
    unit: token
    aggregation: sum
    dimensions:
      - model
      - direction
apis:
  - name: NVIDIA Developer API
    baseURL: https://integrate.api.nvidia.com
    tags:
      - GPU
      - AI
      - Machine Learning
      - Inference
    serviceName: NVIDIA Developer API
    serviceCategory: AI Infrastructure
unitEconomics:
  - name: Cost per inference token (hosted)
    metric: build_credits_cost / inference_tokens
  - name: GPU utilization
    metric: gpu_busy_seconds / gpu_allocated_seconds
  - name: Cost per million tokens (self-hosted)
    metric: (ai_enterprise_subscription + power + amortized_hardware) / (inference_tokens / 1000000)
maintainers:
  - FN: API Evangelist
    email: info@apievangelist.com