# finops-foundation-finops-for-ai-overview-2026-02-17

## Veille

Official FinOps Foundation guide to AI: token economics, KPIs, caching, prompt optimization, model selection, and extension of the FinOps Framework's 14 capabilities to GenAI services - FinOps Foundation

## Titre Article

FinOps for AI Overview

## Date

2026-02-17

## URL

https://www.finops.org/wg/finops-for-ai-overview/

## Keywords

FinOps Foundation, token economics, cost per token, cost per inference, LLM optimization, caching, prompt engineering, model selection, MMLU, Crawl Walk Run, AI FinOps KPIs, showback, unit economics, AI ROI, FinOps for AI certification

## Authors

FinOps Foundation — groupe de travail (Brent Eubanks/Wayfair, James Barney/MetLife, Eric Lam/Google, Adam Richter/AWS, Rahul Kalva/Wells Fargo, JJ Sharma/KPMG, Karl Hayberg/EY, et al.)

## Ton

**Profile**: Standards-body perspective (multi-company working group), normative and exhaustive register, experienced FinOps practitioner level

**Description**: Collective reference document, co-written by a broad working group representing major companies (Google, AWS, MetLife, Wells Fargo, Roche, Accenture, KPMG, EY…). The tone is that of a standard under CC BY 4.0 license: neutral, structured, vendor-agnostic, aiming for exhaustiveness rather than argument. Each KPI is given with its formula and a worked example; each of the 14 FinOps Framework capabilities is screened for "common to cloud / different for AI." Progression is framed by a Crawl-Walk-Run maturity model. Target audience: practicing FinOps professionals, Finance/Engineering/Data teams, and candidates for the *Certified FinOps for AI* certification.

## Pense-betes

- **Token = fundamental unit** of LLM API consumption; *"the meters, or elements of charge, can be very different"* from classic cloud metrics (gap between user input tokens and billed semantic tokens).
- **KPIs with formulas and examples**:
- **Cost Per Token** = Total cost / Tokens used (e.g., $2,500 / 1M = $0.0025).
- **Cost Per Inference** = Inference costs / Number of requests (e.g., $5,000 / 100,000 = $0.05).
- **Training Cost Efficiency** = Training cost / performance metric (e.g., $10,000 / 95% = $105 per point).
- **ROI** = (Benefits − Costs) / Costs × 100 (e.g., 150%).
- **LLM Model Choice Quality Score Alignment**: compares the minimum required MMLU to the MMLU of the model used → identifies waste (e.g., a sentiment analysis task requiring MMLU 54 should not run on GPT-4).
- **Token optimization**: shortening prompts while preserving clarity; **caching** of repeated responses; *"avoid using the most complex and expensive models for every task"*; **model distillation** (smaller versions of GPT-4/Claude for production).
- **14 capabilities of the FinOps Framework** revisited for AI — the most differentiated: Data Ingestion, **Allocation** (traceability of *multi-agent workloads*, absence of a standard framework), Planning & Estimating (estimating successful outputs vs. hallucinations), Forecasting (lower predictability in Crawl/Walk), Budgeting, Benchmarking (per-token, few external benchmarks), Unit Economics (cost-per-call, customer satisfaction per dollar), Rate Optimization (OpenAI Scale Tier), Cloud Sustainability (environmental impact per request).
- **Crawl → Walk → Run maturity model**: Crawl = *fail-fast* prototyping, manual calculations; Walk = basic tracking automation + anomaly detection; Run = advanced tracking & anomaly detection, integrated financial metrics, avoid cutting costs that compromise non-functional requirements.
- **8 pricing models**: on-demand, reserved/CUD, provisioned capacity (OpenAI Scale Tier, Azure PTU), spot/batch, subscription, tiered, free/freemium, hybrid.
- **Showback** favored (visibility without immediate billing) as an awareness tool ahead of chargeback.
- **Tools cited**: Langfuse, Langsmith, Prometheus, Grafana, OTEL; AWS/GCP/Azure native (Bedrock, Vertex AI, Azure OpenAI); GGUF/ONNX/OpenVINO/TensorRT optimization frameworks; vector DBs (Kendra, OpenSearch, pgvector, Cosmos DB).
- **Note**: the document does not yet treat **AI agents** as a distinct category (only *multi-agent workloads* are mentioned under Allocation) — hence the value of vendor complements (Finout, Orq.ai).
- **Ecosystem**: *Certified FinOps for AI* certification; **FinOps X 2026** conference (June 8-11, San Diego); CC BY 4.0 license.
- **Watch link**: official doctrinal foundation of the agentic FinOps cluster — extended by [[finout-finops-ai-agents-four-step-allocation-framework-2026-04-27]] (allocation), [[finout-cpo-guide-llm-rag-agents-agentic-token-multipliers-2025-11-02]] (token multipliers), [[orq-ai-finops-ai-agents-cost-per-outcome-hosseini-2026-04-15]] (cost per outcome). Cross-references Gupta (cost-per-token, marginal token utility) and the **Cost optimization** slot.

## RésuméDe400mots

*FinOps for AI Overview* is the FinOps Foundation's reference guide, co-authored by a broad working group (Google, AWS, MetLife, Wells Fargo, Roche, Accenture, KPMG, EY…) and published under a CC BY 4.0 license. It extends the FinOps discipline to generative AI services, starting from the token as the fundamental unit of consumption, whose "meters" differ profoundly from classic cloud metrics.

The document provides a battery of **KPIs with formulas and worked examples**: Cost Per Token (total cost / tokens), Cost Per Inference (inference costs / requests, e.g., $0.05), Training Cost Efficiency (cost / accuracy point), ROI ((benefits − costs)/costs × 100), and above all the *LLM Model Choice Quality Score Alignment*, which compares the minimum MMLU score a task requires against the MMLU of the model actually used, to detect over-provisioning (a sentiment analysis task requiring MMLU 54 should not run on GPT-4).

On the **optimization** side, the focus is on token reduction (shortening prompts while preserving clarity), **caching** of repeated responses, **model selection** (*"avoid using the most complex and expensive models for every task"*), and **model distillation** for production.

The structural core maps the **14 capabilities of the FinOps Framework** onto "common to cloud" versus "different for AI." The most affected: **Allocation** (traceability of *multi-agent workloads*, absence of a standard framework), Planning (estimating successful outputs and separating them from hallucinations), Forecasting (lower predictability in early phases), Benchmarking (per-token metrics, few external benchmarks), Unit Economics (cost-per-call, customer satisfaction per dollar), and Rate Optimization (volatile pricing such as OpenAI Scale Tier).

Maturity progression follows a **Crawl → Walk → Run** model: *fail-fast* prototyping and manual calculations at the start; basic tracking automation and anomaly detection next; advanced tracking, integrated financial metrics, and vigilance against cutting costs that compromise non-functional requirements in the Run phase. The document lists **eight pricing models** (on-demand, reserved/CUD, provisioned — OpenAI Scale Tier, Azure PTU —, spot/batch, subscription, tiered, freemium, hybrid) and favors **showback** as an awareness lever ahead of chargeback.

A notable limitation: **AI agents** are not yet treated as a distinct category (only *multi-agent workloads* surface under Allocation), which explains the value of vendor complements. The guide is accompanied by a *Certified FinOps for AI* certification and points to FinOps X 2026 (June, San Diego).

## GrapheDeConnaissance

- FinOps Foundation —publie→ FinOps for AI Overview (DOCUMENT, 0.98)
- FinOps for AI Overview —affine→ Framework FinOps (METHODOLOGIE, 0.97)
- token —est_instance_de→ unité fondamentale de consommation LLM (CONCEPT, 0.97)
- Cost Per Token —est_basé_sur→ Coût total / tokens utilisés (CONCEPT, 0.96)
- sélection de modèle —réduit→ coûts inutiles (CONCEPT, 0.94)
- MMLU —permet→ alignement capacité modèle / besoin tâche (CONCEPT, 0.92)
- caching —réduit→ consommation de tokens (CONCEPT, 0.93)
- prompt engineering concis —réduit→ tokens (15-25%) (CONCEPT, 0.88)
- 14 capacités revisitées pour l'IA —fait_partie_de→ Framework FinOps (METHODOLOGIE, 0.95)
- modèle Crawl-Walk-Run —s_applique_à→ maturité FinOps IA (CONCEPT, 0.94)
- FinOps for AI Overview —affirme_que→ l'allocation est compliquée par les multi-agent workloads (AFFIRMATION, 0.9)
- showback —permet→ prise de conscience des coûts (CONCEPT, 0.9)
- FinOps Foundation —a_créé→ certification Certified FinOps for AI (CONCEPT, 0.92)

---
Canonical: https://www.thekb.eu/en/fiches/finops-foundation-finops-for-ai-overview-2026-02-17/