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: Google TensorFlow
providerId: google-tensorflow
publisherName: TensorFlow Authors / Google LLC
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:
  - FinOps
  - FOCUS
  - AI
  - Machine Learning
  - Open Source
description: 'FOCUS-aligned FinOps for Google TensorFlow: open-source, no-cost framework. Cost surfaces
  are entirely the underlying compute, storage, and engineer-time consumed by self-hosted TensorFlow
  Serving and training pipelines, not TensorFlow license fees.'
sources:
  - https://www.tensorflow.org/
  - https://github.com/tensorflow/serving
billingModel:
  pricingCategory: Open Source
  billingFrequency: On-Demand
  billingCurrency: USD
  chargeCategories:
    - Usage
focusColumns:
  ServiceName: Google TensorFlow
  ServiceCategory: AI Infrastructure
  ProviderName: TensorFlow / Google
  PublisherName: Google LLC
  InvoiceIssuerName: 'N/A (open source)'
  BillingCurrency: USD
  ChargeCategory: Usage
meters:
  - name: serving_inference_requests
    description: Inference requests handled by self-hosted TensorFlow Serving (operational metric, not
      a billing line)
    unit: request
    aggregation: sum
    dimensions:
      - model
      - version
      - environment
  - name: serving_compute_hours
    description: Underlying compute hours consumed by TensorFlow Serving processes
    unit: instance-hour
    aggregation: sum
    dimensions:
      - instance_type
      - accelerator
      - region
  - name: training_compute_hours
    description: Compute consumed during model training / fine-tuning
    unit: instance-hour
    aggregation: sum
    dimensions:
      - instance_type
      - accelerator
      - region
  - name: model_artifact_storage
    description: Storage occupied by trained model artifacts and checkpoints
    unit: GB-month
    aggregation: avg
    dimensions:
      - storage_class
      - region
  - name: hub_model_downloads
    description: Pre-trained model downloads from TensorFlow Hub / Kaggle Models
    unit: download
    aggregation: sum
    dimensions:
      - model
principles:
  - name: Visibility
    description: Track inference and training compute / storage cost in the underlying cloud bill (GCP,
      AWS, Azure) — TensorFlow itself produces no invoice. Use Prometheus / OpenTelemetry exporters to
      attribute requests-per-model.
  - name: Allocation
    description: Tag the cloud resources running TensorFlow Serving / training jobs with team, model,
      and environment labels so attached compute spend rolls up to the consuming product.
  - name: Optimization
    description: Use TFLite / quantization / pruning to shrink models; right-size accelerators; enable
      batching in TF Serving; share inference clusters across low-traffic models.
  - name: Accountability
    description: Each ML team owns the cloud cost of training and serving its models; review monthly
      training-run cost vs. model-quality lift in MLOps reviews.
maintainers:
  - FN: Kin Lane
    email: kin@apievangelist.com