AI cost tracking
AI cost tracking is collecting what an organization spends on AI, from model APIs such as Anthropic and OpenAI, cloud AI services such as Amazon Bedrock, and AI tools such as coding assistants, in one place, and attributing it to the teams, products and models behind it. It applies FinOps to AI spend.
In plain English
Think of a household with streaming subscriptions on three cards, a pay-per-view habit and a family plan nobody remembers signing up for. Nobody is overspending on purpose; nobody can see the total either.
Worked example
| Anthropic API: the support assistant and the Data team's agents, by workspace | $5,200 |
|---|---|
| Amazon Bedrock, on the AWS bill: document search | $2,800 |
| OpenAI API: internal tools, by project | $1,900 |
| Cursor: 40 developers' seats and usage | $1,600 |
| AI spend in March | $11,500 |
That's equal to 10% of the $115,000 cloud bill, but only the Bedrock line arrives on it; the rest comes on three separate invoices, each with its own units and IDs. Tracking starts with getting all four into one view, by team and model.
Why AI cost tracking matters
AI spend grows faster and less visibly than most cloud spend. It arrives on several bills, it's priced in tokens and seats rather than hours, and a single prompt change can double it overnight. Without tracking, the first sign is usually an invoice.
The FinOps Foundation applies the same goal to AI as to the rest of technology spend: know what each team, product and model costs, and what it delivers. Its guidance suggests unit metrics such as cost per inference, per user and per successful outcome.
What to track
- Spend by source. Model APIs, cloud AI services and AI tools, in one currency and one view.
- Spend by owner. Workspaces, projects and API keys mapped to teams and products, the same way you allocate cloud cost.
- Spend by model and token type. Which models cost most, and how much is input, cached input and output.
- Unit costs. Cost per request, per conversation or per accepted line of code, so growth can be told apart from waste.
Common mistakes
- Tracking only the API bill. Coding assistants, cloud AI services and AI features in other tools are AI spend too. Miss them and the total is wrong.
- Sharing one API key across teams. A shared key makes attribution impossible. Give each team or product its own workspace, project or key.
- Paying for idle capacity. Provisioned throughput and unused seats bill whether or not anyone uses them. Check their use as you would a commitment.
- Setting no limits. AI spend can jump in a day. Put budgets and anomaly alerts on each AI provider, as on any cloud account.
Where to find it in AWS, Azure and Google Cloud
| Cloud | What it is called | Where to find it |
|---|---|---|
 AWS | Amazon Bedrock | Bedrock is billed on the AWS invoice, so it shows up in Cost Explorer and the Cost and Usage Report alongside other AWS services, and in AWS Budgets and Cost Anomaly Detection. |
 Azure | Azure OpenAI in Microsoft Foundry | Usage is billed to the Azure subscription and appears in Cost Management with other Azure services. Provisioned throughput units bill hourly whether or not they serve requests. |
 Google Cloud | Gemini on Google Cloud | Gemini API and Vertex AI usage is billed to the Cloud Billing account and appears in the BigQuery billing export by project and SKU. |
How CloudYali helps
CloudYali puts Anthropic, OpenAI, Cursor, Amazon Bedrock and Gemini spend next to your cloud bill, broken down by model, workspace, API key and person, with budgets and anomaly alerts on the tools that support them.
Related terms
AI, data & SaaS costs
Cost per token
Cost per token is what a language model charges for each token it reads or writes, usually quoted per million tokens, with input, cached input and output tokens priced separately. Output tokens typically cost several times more than input, so the mix of token types, not just the volume, decides the bill.
Protocols & tooling
Model Context Protocol (MCP)
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. An MCP server exposes tools, resources and prompts over JSON-RPC, and any MCP client, such as Claude, ChatGPT or Cursor, can use them, so one integration works across many AI assistants.
FinOps foundations
Unit economics
Unit economics is measuring cloud cost per unit of business value, such as cost per order, per customer or per API call, by dividing what a product costs by how much it delivered. It shows whether spend is growing because the business is growing or because the product is becoming less efficient.
Allocation & accountability
Cost allocation
Cost allocation is assigning every part of the cloud bill to the team, product or customer responsible for it, using accounts, tags and agreed rules for shared costs. It turns one invoice into numbers each owner can act on, and it is the basis for showback, chargeback and unit economics.
Governance & planning
Budget
A budget is a planned spending limit for a period, such as a team's monthly cloud cost, with alerts when actual or forecasted spend crosses thresholds you set, such as 80% and 100%. On AWS, Azure and Google Cloud a budget warns rather than stops spending, unless you attach actions or automation to it.
Governance & planning
Cost anomaly
A cost anomaly is spend well outside what a service, account or team normally costs, such as a day costing several times its recent average. Anomaly detection compares each day with a baseline learned from history and alerts on the difference, so a runaway job is caught in days, not at the end of the month.
Frequently asked questions
What counts as AI spend?
Model API usage, such as Anthropic or OpenAI; AI services on your cloud bill, such as Amazon Bedrock or Gemini on Google Cloud; and AI tools billed per seat or usage, such as coding assistants.
How do I attribute AI cost to teams?
Give each team or product its own workspace, project or API key, and map those to owners the same way you map cloud accounts and tags.
Which unit metrics work for AI?
Cost per request or conversation for products, cost per accepted line or agent request for coding tools, and cost per million tokens by model to compare rates.
