specification: FinOps Framework
specificationVersion: '1.0'
schema: https://www.finops.org/framework/
provider: vLLM
providerId: vllm
created: '2026-05-08'
# Provenance stamped 2026-08-11: this artifact was written by the API Evangelist
# bulk sweep dated 2026-05-08, not harvested from the provider. See roadmap#35.
method: generated
modified: '2026-05-08'
reconciled: true
tags:
- LLM
- Inference
- Open Source
- GPU
- OpenAI Compatible
- Self-Hosted
- FinOps
- Cost Management
- FOCUS
description: >-
  FOCUS-aligned FinOps profile for vLLM. The software is free. Cost is GPU compute and storage on your
  chosen platform (cloud spot/on-demand GPUs, on-prem hardware, or a managed-vLLM provider). Optimize
  via continuous batching (default in vLLM), prefix caching, paged-attention sizing, model
  quantization (AWQ/GPTQ/INT8/FP8), and routing across instance sizes.
notes: Cost is GPU infra; vLLM provides batching/quantization levers to reduce $/token.
sources:
- https://docs.vllm.ai/
- 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: vLLM Project
serviceCategory: LLM Inference
billingModel:
  pricingCategory: Free OSS / Compute
  billingFrequency: N/A (project) / Hourly (infra)
  billingCurrency: USD
  chargeCategories:
  - Usage
focusColumns:
  ServiceName: vLLM
  ServiceCategory: LLM Inference
  ProviderName: vLLM Project
  PublisherName: vLLM Project
  InvoiceIssuerName: N/A
  BillingCurrency: USD
  ChargeCategory: Usage
meters:
- name: gpu_hours
  description: GPU hours consumed by vLLM serving processes.
  unit: hour
  aggregation: sum
  dimensions:
  - gpu_type
  - model
- name: tokens_served
  description: Output tokens generated (operational meter for $/token tracking).
  unit: token
  aggregation: sum
  dimensions:
  - model
principles:
- name: Visibility
  description: Pull GPU utilization and tokens-served metrics; compute effective $/1k tokens.
- name: Allocation
  description: Allocate GPU hours to consuming AI features/teams.
- name: Optimization
  description: Use continuous batching, prefix caching, and quantization; right-size GPUs per workload.
- name: Accountability
  description: Infra/ML platform team owns vLLM deployment and GPU spend.
maintainers:
- FN: Kin Lane
  email: kin@apievangelist.com