이 페이지는 아직 한국어로 제공되지 않습니다. 번역 작업 중입니다.
현재 번역 프로젝트에 대한 질문이나 피드백이 있으신 경우 언제든지 연락주시기 바랍니다.

Overview

AI Costs in Cloud Cost Management gives FinOps and engineering teams a unified destination for analyzing AI spend across providers, including Amazon Bedrock, Amazon SageMaker, Anthropic, Azure Foundry, Cursor, Databricks Mosaic AI, GitHub Copilot, Google Gemini, OpenAI, Snowflake AI, and Vertex AI. View total AI spend alongside your existing cloud infrastructure costs, analyze it with normalized tags, track cost anomalies, identify optimization opportunities, and attribute usage to the specific users and API keys driving it.

Prerequisites

To use AI Costs, you must have at least one of the following supported providers set up for Cloud Cost Management:

AI providerSetup method
Amazon BedrockAWS integration
Amazon SageMakerAWS integration
Anthropic (Enterprise and Platform)SaaS integration
Azure Foundry (Models and Tools)Azure integration
Google GeminiGoogle Cloud integration
OpenAI (Platform only)SaaS integration
Vertex AIGoogle Cloud integration
GitHub CopilotGitHub Copilot
CursorCursor
Databricks Mosaic AISaaS integration
Snowflake AISaaS integration

AI cost summary

After connecting your AI providers, navigate to Cloud Cost > Summarize > AI to view the AI cost summary page.

The AI cost summary dashboard, showing daily spend trends over a one-month period, a list of top spend drivers, and an anomaly graph.

The AI cost summary page provides:

  • Total AI Cost: Aggregated AI cost and cost change over the selected time frame.
  • Daily AI Cost: Daily cost trends across the selected providers over the selected time frame. Use the Filter to dropdown to define which providers appear in the graph.
  • Top Cost Drivers: The models, projects, services, and users generating the most spend.
  • Active AI Cost Anomalies: Cost anomalies surfaced proactively across all connected providers. Select an anomaly to open a side panel with more details and options for further action.
  • AI Cost Recommendations: Cost recommendations and optimization opportunities surfaced across all connected providers. Select a recommendation to open a side panel with more details and options for further action.
  • AI Cost Dashboards: Out-of-the-box dashboard templates for each supported provider, combining cost data with usage signals such as token consumption, model distribution, and user analytics.

Normalized AI tags

AI cost data from all supported providers is normalized to a consistent set of tags. Use these tags to filter, group, compare, and plan AI spend across dashboards, monitors, budgets, forecasts, and other Datadog tools. Use Cloud Cost Explorer to query and compare spend across providers without writing per-provider logic.

The following tags are available for all supported AI providers:

Tag name                                   Tag description
providernameThe AI provider.
ai_vendorThe AI provider and plan that the cost belongs to (for example, amazon_bedrock, anthropic_enterprise, or anthropic_platform).
modelThe AI model identifier (for example, claude-opus-4-6, gpt-4.1).
model_nameThe human-readable model name (for example, Claude Opus 4.6).
token_directionWhether tokens are being consumed (input) or generated (output) within a service or application.
token_categoryThe specific category of tokens consumed, such as input tokens, output tokens, or tokens related to caching and search operations (for example, cached input, cache write, standard input, output).
projectThe project, workspace, or environment the AI costs belong to.

Attribute AI spend to sources

Each provider reports its own tags that attribute AI spend to the users, API keys, projects, and other sources that generated it. For Anthropic and OpenAI, Datadog also adds tags through out-of-the-box (OOTB) allocation rules, which require no configuration.

Provider-specific tags

In addition to the normalized AI tags, each provider reports its own set of tags. The AI cost summary page uses these tags to break down, filter, and group each provider’s spend. The label column shows the name the tag uses on the AI cost summary page.

TagLabelDescription
model_nameModelThe model that generated the cost.
projectProjectThe project that the cost belongs to.
aws_iam_principal_userIAM UserThe IAM user that made the request.
aws_iam_principal_session_nameIAM Session NameThe session name of the IAM role session that made the request.
aws_iam_principal_application_roleIAM Application RoleThe IAM role that the application assumed to make the request.

The IAM principal tags are available only when your AWS Cost and Usage Report (CUR 2.0) includes IAM principal data.

TagLabelDescription
sagemaker_resource_typeResource TypeThe type of SageMaker resource that generated the cost (for example, app).
sagemaker_resource_nameResource NameThe name of the SageMaker resource that generated the cost.
sagemaker_app_typeApp TypeThe type of SageMaker app. Available only for resources where sagemaker_resource_type is app.
sagemaker_app_nameApp NameThe name of the SageMaker app. Available only for resources where sagemaker_resource_type is app.
sagemaker_componentComponentThe SageMaker component that generated the cost.

Anthropic reports a different set of tags for each plan.

Anthropic Enterprise

TagLabelDescription
model_nameModelThe model that generated the cost.
user_emailUserThe email address of the user who generated the cost.
productProductThe Anthropic product that generated the cost (for example, claude_code).
context_windowContext WindowThe context window size range of the request (for example, tier_0-200k or tier_200k-1m).
speedSpeedThe output speed mode of the request.

Anthropic Platform

TagLabelDescription
model_nameModelThe model that generated the cost.
api_key_nameAPI Key NameThe name of the API key that made the request.
workspace_nameWorkspaceThe Anthropic workspace that the cost belongs to.
context_windowContext WindowThe context window size range of the request (for example, tier_0-200k or tier_200k-1m).
service_tierService TierThe service tier that processed the request (for example, Standard).
cost_typeCost TypeThe type of charge (for example, tokens).

Additional tags

The following Anthropic tags aren’t shown on the AI cost summary page, but are available for filtering and grouping across Cloud Cost Management:

TagDescription
api_key_idThe unique identifier of the API key that made the request.
modelAn alias for model_id, provided for consistency with usage metrics.
model_idThe canonical Anthropic model identifier (for example, claude-opus-4-6).
org_idThe unique identifier of the Anthropic organization.
org_nameA tag-normalized version of the Anthropic organization’s name.
user_idThe unique identifier of the user who generated the cost.
user_nameThe name of the user who generated the cost.
workspace_idThe unique identifier of the Anthropic workspace.

Azure Foundry spend is split into Foundry Models, which covers the Foundry Models meter category, and Foundry Tools, which covers all other AI meter categories.

Azure Foundry Models

TagLabelDescription
model_nameModelThe model that generated the cost.
resourcegroupnameResource GroupThe Azure resource group that the cost belongs to.
metercategoryCategoryThe Azure meter category of the charge (for example, Foundry Models).
metersubcategorySubcategoryThe Azure meter subcategory of the charge.

Azure Foundry Tools

TagLabelDescription
resourcegroupnameResource GroupThe Azure resource group that the cost belongs to.
metercategoryCategoryThe Azure meter category of the charge (for example, Foundry Tools or Cognitive Services).
metersubcategorySubcategoryThe Azure meter subcategory of the charge.
TagLabelDescription
model_nameModelThe model that generated the cost.
user_emailUserThe email address of the user who generated the cost.
max_modeMax ModeWhether the request used Cursor’s Max Mode.
is_headlessHeadless ModeWhether the request ran in headless mode, without the Cursor editor.
billing_groupBilling GroupThe Cursor billing group that the user belongs to.
TagLabelDescription
billing_origin_productBilling Origin ProductThe Databricks product that generated the cost (for example, MODEL_SERVING or VECTOR_SEARCH).
model_nameModelThe model that generated the cost.
workspace_nameWorkspaceThe Databricks workspace that the cost belongs to.
TagLabelDescription
model_nameModelThe model that generated the cost.
projectProjectThe Google Cloud project that the cost belongs to.
TagLabelDescription
organization_nameOrganizationThe GitHub organization that the cost belongs to.
repository_nameRepositoryThe GitHub repository that the cost belongs to.
TagLabelDescription
model_nameModelThe model that generated the cost.
user_emailUserThe email address of the user who generated the cost.
projectProjectThe OpenAI project that the cost belongs to.
api_key_nameAPI Key NameThe name of the API key that made the request.

Additional tags

The following OpenAI tags aren’t shown on the AI cost summary page, but are available for filtering and grouping across Cloud Cost Management:

TagDescription
account_idThe unique identifier of the OpenAI account.
account_nameThe name of the OpenAI account.
api_key_idThe unique identifier of the API key that made the request.
batchWhether the request used the OpenAI Batch API.
endpointThe OpenAI API endpoint that the request was sent to.
modelThe OpenAI model identifier (for example, gpt-4.1).
org_idThe unique identifier of the OpenAI organization.
project_idThe unique identifier of the OpenAI project.
project_nameThe name of the OpenAI project.
user_idThe unique identifier of the user who generated the cost.
TagLabelDescription
account_nameAccountThe Snowflake account that the cost belongs to.
TagLabelDescription
model_nameModelThe model that generated the cost.
projectProjectThe Google Cloud project that the cost belongs to.

Map tags to teams

Configure Tag Pipelines to map provider-specific tags (such as user_email) to teams, services, or business units for aggregate reporting:

The Tag Pipelines Rule Setup page, showing user_email values mapped to team values through an existing reference table, and additional tag mapping options.

After mapping, attributed spend appears in provider-specific dashboards and Cost Reports:

A provider-specific dashboard with a stacked bar graph showing daily provider spend attributed by team and model name, and a summary list of the spend attributions.

Allocate shared AI gateway spend

If your teams send LLM requests through a shared AI gateway, you can split that spend across the teams, users, or services that sent the requests. For more information, see AI Gateway Allocation.

Further reading