specification: FinOps Framework
specificationVersion: '1.0'
schema: https://www.finops.org/framework/
provider: Predibase
providerId: predibase
created: '2026-06-20'
modified: '2026-06-20'
reconciled: true
tags:
- AI
- LLM
- Fine-Tuning
- Inference
- LoRA
- FinOps
- Cost Management
- FOCUS
description: >-
  FinOps view of Predibase spend. Predibase bills three usage meters:
  serverless inference per token (scaled by model size), batch inference at a
  flat per-million-token rate, fine-tuning (training) per token of training data
  scaled by base-model size, and dedicated deployments per GPU-hour by
  accelerator. LoRAX multi-LoRA serving packs many adapters onto one GPU,
  reducing dedicated serving cost versus one deployment per fine-tuned model.
notes: >-
  Per-model and per-accelerator rates are subject to change; verify against the
  Predibase pricing page during reconciliation. Predibase was acquired by Rubrik
  in June 2025.
sources:
- https://predibase.com/pricing
- https://docs.predibase.com
- https://focus.finops.org/focus-specification/v1-3/
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/
publisherName: Predibase
serviceCategory: AI and Machine Learning
billingModel:
  pricingCategory: Usage-Based
  billingFrequency: Monthly
  billingCurrency: USD
  chargeCategories:
  - Usage
  - Purchase
  - Adjustment
focusColumns:
  ServiceName: Predibase
  ServiceCategory: AI and Machine Learning
  ProviderName: Predibase
  PublisherName: Predibase
  InvoiceIssuerName: Predibase
  BillingCurrency: USD
  ChargeCategory: Usage
  PricingCategory: Usage-Based
meters:
- name: serverless_inference_tokens
  description: Tokens served via serverless / shared endpoints, billed per 1M tokens scaled by model size.
  unit: tokens
  aggregation: sum
  dimensions:
  - account
  - model
  - deployment
- name: batch_inference_tokens
  description: Tokens served via batch inference jobs, billed at a flat per-1M-token rate.
  unit: tokens
  aggregation: sum
  dimensions:
  - account
  - model
- name: finetuning_training_tokens
  description: Training tokens consumed by supervised / GRPO fine-tuning jobs, billed per 1M tokens scaled by base-model size.
  unit: tokens
  aggregation: sum
  dimensions:
  - account
  - base_model
  - task
- name: dedicated_gpu_hours
  description: GPU-hours consumed by dedicated / private deployments, billed per accelerator type (A10, A100, H100).
  unit: hours
  aggregation: sum
  dimensions:
  - account
  - deployment
  - accelerator
principles:
- name: Visibility
  description: Pull Predibase usage exports; inspect token burn per deployment and GPU-hours per accelerator.
- name: Allocation
  description: Tag API tokens and deployments per workload/team; map adapters and repos to internal cost centers.
- name: Optimization
  description: Prefer serverless / batch for spiky or non-realtime traffic; consolidate many fine-tuned adapters onto one LoRAX-enabled deployment instead of one deployment per model; right-size the GPU accelerator; pick smaller base models when sufficient.
- name: Accountability
  description: Assign owners per project/deployment; review token and GPU-hour spend monthly against budget.
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com