FinOps for AI Overview

At a Glance

AI is core FinOps work. Most of the practice carries over; tokens, volatile pricing, and new stakeholders are what change.
Map spend across categories. Model APIs, cloud services, SaaS and developer tools, GPUs, and data centers can all feed one AI Scope.
See it before you optimize it. Allocate first, then right-size models, use tokens efficiently, and commit once usage is stable.
Measure the use case, not the token. Track cost per successful outcome, and report it as a range.
Bring cost to the value conversation. Pair allocated cost with outcomes the business owns to show what AI returns.

New AI Cost and Usage Challenges, Same FinOps Approach
Fundamentals of AI Cost and Usage
Measuring AI’s Business Value
Best Practices by FinOps Domain
Building Incrementally: Crawl, Walk, Run
KPIs and Metrics
Regulatory and Compliance Considerations
How AI Changes Each FinOps Capability
Go Deeper
Acknowledgments

New AI Cost and Usage Challenges, Same FinOps Approach

Organizations across industries and maturity levels use AI, especially generative AI and large language models (LLMs), to improve products, raise productivity, and serve customers. For FinOps teams, AI is no longer an emerging topic. According to the State of FinOps 2026, 98 percent of respondents manage AI spend, up from 31 percent two years earlier.

Some AI challenges are the same as those of any new technology: new terminology, new stakeholders, and new pricing and discount models to learn. Others are specific to AI. Pricing units such as tokens are unfamiliar, costs change quickly, purchasing channels multiply, and people far outside IT (Product, marketing, sales, and Leadership) now drive AI spend directly.

This FinOps practitioner’s overview of AI cost and usage explains how AI spend works, applies the FinOps Framework to it, and shows how FinOps data feeds the conversation about AI value. It complements the FinOps for AI Technology Category guidance in the Framework.

Scope note: This Paper covers FinOps for AI, managing the cost and value of AI. Using AI to perform FinOps is a separate topic covered under AI for FinOps.

The Tokenomics Foundation, part of the Linux Foundation’s technology value umbrella alongside the FinOps Foundation, develops open methods for the full economics of AI, from the energy and capital that produce it to the value it creates. This Paper links to that work where it helps practitioners go deeper. For a starting definition, see What Is AI Tokenomics?

Fundamentals of AI Cost and Usage

What Stays the Same

AI services can seem entirely new, but much of an existing FinOps practice applies directly.

The basic equation holds. Cost equals price times quantity, so practitioners can lower cost by improving rates or by using less.
Cloud AI costs appear in cloud billing data. AI services bought from cloud providers show up alongside other cloud costs, often in FinOps Open Cost and Usage Specification (FOCUS) format.
Tagging and allocation still work for many services, with adjustments for shared environments, training, and API-based usage.
Commitments still apply. Many AI resources are eligible for commitment discounts or reserved capacity, and rate management processes are similar.

What Is Different

New units of cost. Tokens, API calls, and provisioned throughput units behave differently from instance hours. The tokens a user types are not always the tokens billed, because applications add system prompts, history, and retrieved content. See GenAI FinOps: How Token Pricing Really Works.
Volatile pricing. Models are released and repriced often, in many versions and tiers, and prices can move sharply in either direction.
Thin allocation data. New SKUs and direct model provider bills often lack native tagging, so allocation needs engineering instrumentation.
Capacity constraints. GPU capacity and provisioned model throughput can be scarce in some regions, which affects availability, commitments, and price.
Quality is part of cost. The right model is the least expensive one that meets the quality bar for the task. Choosing a model is a cost decision and a quality decision at once.
New stakeholders. Teams with little experience of FinOps, or of being accountable for spend, build and buy AI directly.

Where AI Spend Shows Up

AI spend crosses FinOps Technology Categories, so a FinOps Scope for AI often draws from several at once. Common channels include:

Channel	Typical Cost Unit	Technology Category
Direct model provider APIs	Per token	AI (token-based SaaS)
Cloud provider model services and AI platforms	Per token, per call, or per hour, on the cloud bill	Public Cloud
Model routers and aggregators	Credits or per token	AI (token-based SaaS)
Embedded AI in SaaS products and developer tools	Per seat, add-on, or usage	SaaS
Cloud or specialized GPU providers for training and self-hosted models	GPU hours, storage, networking	Public Cloud
Owned or colocated AI infrastructure	Capital, power, and operating cost	Data Center

Each channel has its own visibility, allocation, and optimization characteristics. For direct model provider APIs, the hardest channel to manage, see Managing AI Token Costs. For choosing between building, buying, and hosting, see Choosing an AI Approach and Infrastructure Strategy.

Pricing Models

AI pricing mixes cloud-style and SaaS-style models, often within one Scope.

Model	How It Charges	Best Fit	Watch For
On-demand	Per token, call, or hour used	Variable or experimental workloads	Runaway usage without monitoring and limits
Commitment discounts	Discounted rates for committed usage or spend	Predictable training or inference	Underused commitments when workloads change
Provisioned capacity	Fixed block of throughput, billed whether used or not	Latency-sensitive production workloads	Low utilization; may cost more per token than on-demand
Spot, preemptible, and batch	Discounted, interruptible, or asynchronous	Training, retraining, batch inference	Interruptions; needs scheduling and checkpointing
Subscription or seat	Recurring fee per user or platform	Steady use of AI platforms, SaaS AI features, developer tools	Paying for seats that are not used
Tiered or volume	Unit price falls as volume rises	Growing, high-volume workloads	Forecasting errors around tier boundaries
Free tier or preview	Free or discounted for limited use	Experimentation	Cost jumps when previews end or usage scales

For provisioned capacity and when it pays off, see Navigating GenAI Capacity Options.

Who Is Involved

The core FinOps Personas (FinOps Practitioner, Engineering, Finance, Product, Procurement, and Leadership) all apply to AI. What changes is who fills them. Data scientists, machine learning engineers, and employees in non-technical roles who build AI applications act in the Engineering Persona. Product owners of AI tools may sit outside IT. Watch what people do, not their titles, to decide which Persona they are filling, and expect to spend extra time on education with people new to FinOps.

The Tokenomics Foundation’s Personas and Operating Model Map takes an AI-first view of the same question and adds roles such as the AI end user. Reconciling it with the FinOps Personas is ongoing work.

Measuring AI’s Business Value

Many organizations are enthusiastic about AI but unsure how to show what it returns. In the September 2026 State of Tokenomics survey, proving value or ROI was the most commonly named challenge (43 percent). As AI moves from experiments to production, leaders expect a defensible answer.

Value shows up in more than cost savings. It can appear as operational resilience, better user experience, higher productivity, faster time to market, more efficient resource use, and business growth through new revenue or products. A useful value case looks across these dimensions, not only at spend reduced.

FinOps cannot answer the value question alone, but it supplies half of it: trustworthy, allocated cost. To make that contribution useful:

Allocate at the grain the business measures. Cost and value can be compared only when they share a boundary, such as a use case, product, team, or feature.
Include the full cost. Tokens are one part. Add the surrounding infrastructure, such as data stores, orchestration, and observability, and recognize broader costs such as platform labor. The Tokenomics Foundation’s Total Cost of AI work defines that wider boundary.
Tie unit costs to outcomes someone owns. A resolved ticket, a merged pull request, or a processed claim gives the business something to measure against.
Bring in Finance and the outcome owner early. FinOps supplies the cost; the business owns the value claim.

Matching the model to the task matters as much as measuring it. A weak foundation of data cannot support an accurate model, and a frontier model for a simple task is like building a skyscraper when a house will do. Select the least expensive model that meets the required accuracy, latency, and reliability for each use case, and revisit the choice as models improve.

For a method on the outcome side, including how to set a quality floor and avoid double counting, see the Tokenomics Foundation’s Classifying and Measuring Realized Value from AI. For connecting AI investment to business strategy, see the Executive Strategy Alignment Capability and Unlocking AI Business Value with FinOps.

Best Practices by FinOps Domain

The practices below are organized by the four FinOps Domains. Not every practice applies to every organization; choose based on your Scope and maturity level.

Understand Usage & Cost

Ingest every source. Bring together cloud billing data, model provider usage exports, SaaS invoices, and observability data for AI services. Use FOCUS wherever providers support it. FOCUS v1.5, scheduled for ratification in December 2026, is set to add model identity, input and output labels, cache status, and requester attribution to cost data. See the Tokenomics Foundation’s What FOCUS 1.5 Does for AI Cost.

Allocate with tags and keys. Tag training separately from inference, and production separately from experimentation. For API-based services, map each API key to one team, application, and cost center, and use provider attribution features or a proxy where tags are not supported. Example tag keys:

Tag Key	Example Values
Project	Customer_Chatbot, Document_Summarization
Environment	Development, Testing, Production
Workload	Model_Training, Model_Inference, Batch_Inference
Team	Data_Science, ML_Engineering, Platform
CostCenter	AI_Research, Product_AI
Purpose	Experimentation, RealTime_Inference, Batch_Processing
Criticality	High, Medium, Low
ShutdownEligible	True, False

Report through showback first. Show teams their AI costs, broken down by project and workload, before charging them back through Invoicing & Chargeback. Visibility alone often changes behavior. Make reports accessible to the wider set of stakeholders AI brings in.

Detect anomalies early. AI anomalies carry more risk and happen more often than in typical cloud spend. Watch for runaway training jobs, agent loops, prompt changes that double token use, and model switches that change price. Monitor tokens per minute at the application level, not only daily spend. The Tokenomics Foundation’s Big-T Notation helps teams describe how consumption grows with model calls per request and agent depth, which makes unexpected growth easier to spot.

Quantify Business Value

Set two baselines. Establish a cost baseline from current invoices and usage. Then set a functional baseline: the response time, accuracy, reliability, and request volume each AI system needs. Cost targets mean little without the quality they must preserve.

Estimate before building. Estimate the cost of new AI workloads, including the share of outputs that will not meet the quality bar. See Cost Estimation of AI Workloads.

Forecast and budget with Finance. AI forecasts are less predictable than cloud forecasts, especially early on. Adapt Forecasting with shorter windows and more frequent revisions, and use Budgeting to agree with Finance on flexible funding for pilots. See How to Forecast AI Services Costs in Cloud.

Build unit economics. Track cost per inference, per user, and per successful outcome for each use case, and report them as distributions, since AI costs vary from run to run. See KPIs and Metrics.

Optimize Usage & Cost

The Tokenomics Foundation’s Seven-Layer Tokenomics Stack is a useful map of where optimization happens, from energy and hardware at the bottom to model selection, routing, and demand at the top. Organizations that buy model APIs act mostly in the upper layers; those that run their own infrastructure act in all of them.

Right-size the model. Define quality criteria for each task, test candidate models, and choose the least expensive one that meets them. Model routing applies this automatically by sending simple requests to small models. Smaller models can also be produced through distillation, quantization, or pruning.

Use tokens efficiently. Trim system prompts, control output length, manage conversation history, and retrieve only relevant content. Use prompt caching for stable prompt prefixes and batch processing for work that is not time sensitive. The Tokenomics Foundation’s Prompt Cache Explainer shows how prompt structure decides whether caching saves money.

Use infrastructure efficiently. For training and self-hosted models:

Autoscale GPU capacity with demand, and shut down idle instances outside working hours.
Rightsize instances, and use CPUs for experiments and light models that do not need GPUs.
Use spot or preemptible capacity, with checkpointing, for training and batch inference.
Consider accelerators other than GPUs where models support them, weighing the effort to move away from CUDA-based tooling.
Use inference optimization frameworks and serverless options for sporadic workloads.
Tier training data storage by access pattern, and keep data and compute in the same region to avoid transfer fees.

Optimize rates once usage is understood. Use commitment discounts and capacity reservations for predictable workloads, and model the break-even utilization before buying provisioned capacity. Review provider offers often, because AI pricing changes quickly.

Manage the model lifecycle. Add cost and quality checks to deployment pipelines. Retrain when data drift or performance decline justifies it, not on a fixed schedule. Promote a retrained model only when its gains justify its cost, and retire unused models and their storage.

Manage the FinOps Practice

Educate broadly. Teach FinOps basics to AI builders and AI cost basics to Finance, Procurement, and Leadership. See FinOps Education & Enablement and the FinOps for AI Value certification.

Name an owner. AI spend sits between Finance, Engineering, Security, and Procurement, so it needs a named, accountable owner. In the State of Tokenomics survey, organizations with defined ownership of AI economics were 3.7 times more likely to show value to the CFO. Many organizations add an AI investment council or equivalent forum to approve and track AI projects.

Govern without blocking. Through Governance, Policy & Risk, set approved model lists, quotas, rate limits, and budget thresholds, and require review before deploying autonomous agents. Allow lighter rules for pilots so governance does not push teams to work around it.

Choose tools deliberately. Provider dashboards, proxies and gateways, LLM observability platforms, and FinOps platforms each solve part of the problem. See the FinOps Landscape and FinOps for AI: Tools & Services Considerations, and, to build your own tracker, How to Build a Generative AI Cost and Usage Tracker.

Work with Intersecting Disciplines. Coordinate with Security on data sent to models, with IT asset management (ITAM) and Procurement on licenses and contracts, and with Sustainability teams on energy use. See Intersecting Disciplines.

Building Incrementally: Crawl, Walk, Run

AI initiatives carry more uncertainty than most technology projects, so the right approach to cost changes as they mature. Use the FinOps Maturity Model to match cost management to each maturity level.

Maturity Level	Typical Activities	Cost Management Approach	Tracking and Tools
Crawl	Learning, prototypes, proofs of concept, and pilots that validate use cases	Fail fast. Set cost and time limits up front. Spend on what you are validating, such as accuracy, and keep other costs, such as high availability, to a minimum.	Manual tracking; frequently revised budgets; non-financial measures, such as hypotheses confirmed, carry much of the weight
Walk	AI integrated into business processes and producing regular results	Fund what production use requires and no more. Control integration costs, and separate run budgets from release budgets.	Basic automated cost tracking and anomaly detection; financial metrics gain importance; budgets revised less often
Run	AI powers core business processes	Keep costs proportionate to the value delivered. Cut costs that have no effect first, protect agreed service levels, and make other tradeoffs through explicit architecture decisions.	Automated tracking; advanced anomaly detection; unit economics and ROI in leadership reporting; stable budgets by component

KPIs and Metrics

AI systems use many of the same KPIs as other workloads, plus measures specific to AI. The examples below are illustrative. For definitions across Technology Categories, see the FinOps KPI Library and KPIs & Benchmarking.

KPI	What It Measures	Formula	Example
Cost per inference	Operating cost of one model request	Total inference cost / Number of inference requests	$5,000 / 100,000 = $0.05
Cost per million tokens	Effective token rate across models and providers	Total token cost / (Tokens used / 1,000,000)	$2,500 / 500 = $5.00 per million tokens
Cost per successful outcome	Cost of results that meet the quality bar	Total cost of all runs / Runs that meet the quality bar	$1,200 / 8,000 = $0.15
Cost per API call	Average cost of calls to managed AI services	Total API cost / Number of API calls	$1,200 / 240,000 = $0.005
Training cost efficiency	Training cost relative to model performance	Training cost / Performance metric	$10,000 / 95 accuracy points = about $105 per point
Resource utilization efficiency	Share of provisioned GPU or accelerator capacity used	Used capacity / Provisioned capacity	800 / 1,000 GPU hours = 80%
Cache hit rate	Share of input tokens served from cache	Cache reads / (Cache reads + Cache writes + Uncached input)	6M / (6M + 1M + 3M) = 60%
Anomaly cost percentage	Share of AI spend caused by anomalies	Cost of anomalous spikes / Total AI spend	$3,000 / $100,000 = 3%
Return on investment	Financial return relative to cost	(Financial benefits − Costs) / Costs × 100	($50,000 − $20,000) / $20,000 × 100 = 150%
Time to value	Time from starting an initiative to reaching its agreed measurable benefit	Date benefit reached − Start date	January 1 to June 1 = 5 months
Time to first prompt	Time from starting development to first production use	Deployment date − Development start date	January 1 to April 1 = 3 months
Model fit	Share of requests sent to a more capable, more expensive model than the task needs	Requests on overqualified models / Total requests	12,000 / 60,000 = 20%

Two cautions apply. ROI is only as good as its inputs: decide which costs and benefits to include before calculating, and include more than token cost. Model fit depends on a quality bar for each task, set through evaluations Engineering and Product agree on.

Regulatory and Compliance Considerations

Regulation shapes AI cost through the controls, storage, audits, and certifications it requires. The considerations below focus on cost; work with Legal and compliance teams on requirements.

Area	Cost Implications	FinOps Actions
Data privacy (for example, GDPR and CCPA)	Encryption, masking, residency controls, and monitoring add cost; fines add risk. Protecting personal data can also reduce output quality or require costlier solutions.	Tag resources that process sensitive data; account for data residency in workload placement.
Intellectual property and licensing	License fees for proprietary models and datasets; legal exposure from misuse.	Track license terms and costs; involve Legal and Procurement in AI contracts.
Bias and ethics	Audits, retraining, and bias-detection tools.	Budget for periodic audits and remediation.
Sector-specific rules (for example, healthcare and financial services)	Required encryption, audit trails, certifications, and regional or government-specific environments.	Map AI workloads to requirements with compliance teams; budget for certifications and audits.
Data retention	Long-term storage of training and inference data.	Use archive storage tiers; tag datasets with retention policies; review stored data regularly.
Environmental reporting	Carbon reporting and energy efficiency requirements; energy-efficient hardware or offsets.	Use provider carbon data; consider the carbon intensity of regions; see Sustainability.
AI-specific regulation (for example, the EU AI Act)	Risk classification, documentation, and assessments, with stricter requirements for higher-risk uses.	Monitor regulations in key markets; budget for risk assessment and model documentation.

How AI Changes Each FinOps Capability

For each FinOps Capability, the core processes, stakeholders, and principles carry over to AI. The table below summarizes what is different. For full guidance, see FinOps for AI in the Framework.

Domain	Capability	What Is Different for AI
Understand Usage & Cost	Data Ingestion	More sources and third-party providers, with less consistent data quality
	Allocation	Hard to trace which application, feature, or agent consumed model output; no accepted method for multi-agent workloads
	Reporting & Analytics	AI-specific metrics, links to business outputs, and a wider audience
	Anomaly Management	More frequent and riskier anomalies; normal ranges are harder to define
Quantify Business Value	Planning & Estimating	Must estimate unsuccessful outputs and quality thresholds, with few benchmarks
	Forecasting	Lower predictability, mixed pricing units, and more frequent revisions
	Budgeting	Top-down budgets are less accurate; pilots need flexible funding; ownership of AI P&L is often unclear
	KPIs & Benchmarking	New per-token and per-outcome metrics, with few external benchmarks
	Unit Economics	Token-based units plus workload and value metrics, such as cost per resolved case
Optimize Usage & Cost	Architecting & Workload Placement	New steps such as training and fine-tuning; many AI-specific services to combine; placement across channels
	Usage Optimization	Model selection, token efficiency, and caching join instance rightsizing; monitoring is more frequent
	Rate Optimization	Capacity scarcity, provisioned throughput, and fast-changing vendor offers
	Licensing & SaaS	More vendors, AI add-ons to existing contracts, and more dynamic subscriptions
	Sustainability	Energy and carbon per request; idle reserved capacity
Manage the FinOps Practice	Executive Strategy Alignment	AI investment decisions reach the board; projects must be compared and prioritized quickly
	FinOps Practice Operations	New stakeholders unfamiliar with FinOps, across decentralized teams
	FinOps Education & Enablement	New deployment options and pricing models to teach; more facilitation
	FinOps Assessment	AI cost management often matures at a different pace than the rest of the practice, so assess it on its own
	Governance, Policy & Risk	Wider use of quotas, throttles, and model allowlists; lighter rules for pilots
	Invoicing & Chargeback	New invoice formats, provider credits, and developer card payments to reconcile; chargeback for shared models and platforms
	Automation, Tools, & Services	Fast-changing tools that track individual architecture components
	Intersecting Disciplines	Closer work with Security on data sent to models and with Procurement and ITAM on new vendors

FinOps Foundation

Tokenomics Foundation

Acknowledgments

We’d like to thank the following people for their work on this Paper:

Max Audet

COVEO

Joshua Collier

Superhuman fka Grammarly

Ermanno Attardo

Trilogy

Jean Latiere

OptimNow

Borja Martinez

NTT Data

Ilia Semenov

Stealth

Janine Pickard-Green

MagicOrange

Haritza Zubillaga

Roche

Rahul Kalva

Wells Fargo

JJ Sharma

KPMG

Andrew Qu

Everest Systems

Thiago Mozart

RealCloud

Bhaskar Bhowmik

Accenture

Dilli B

Munich Re

Andy Paneof

Adam Richter

Amazon Web Services

Karl Hayberg

We’d also like to thank our supporters and the FinOps Foundation staff for their help: Vas Markanastasakis, Rob Martin, Andrew Nhem, and Samantha White.

Join as an Individual

Join as an Enterprise