Kong measures and controls AI spend while it happens, and your FinOps platform settles it against the invoices once the month closes. Between them you can answer the question finance and engineering both care about, which is what a business outcome costs in AI.

• - Every request is metered and priced as it passes through the gateway, so a runaway agent shows up the day it starts instead of weeks later on the invoice.
• - Usage from OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, and Azure OpenAI is normalized into one schema. Input, output, and cached tokens are counted the same way and priced at your negotiated rates, which makes it easy to compare two models on the same workload.
• - Attribution goes down to the user and the agent. Each request carries its SSO user or gateway key, its cost-center tags, and the agent or app that sent it, and none of that depends on tagging in the clouds.
• - Budget controls sit in the request path. Budgets alert their owners at the thresholds you set and show where spend will land at the current pace, and the AI Rate Limiting Advanced policy can cap tokens or dollars per consumer, returning HTTP 429 once a consumer goes over.
• - Together these give you cost per outcome. Records name the agent, plus the agent session when the client sends one (Claude Code, Codex, and OpenCode do), so you can join them to your own business data and see what each resolved ticket or merged pull request cost.

• - Compute, storage, and the invoices from your cloud and model providers come in through the platform's native connectors.
• - A monthly true-up adjusts Kong's estimates to the invoices, so chargeback matches what you actually paid.
• - Each business unit gets one showback and chargeback view, with its AI spend next to the rest of its cloud bill.

*A team lead asks: “My team spent $1,000 on AI this month. Who spent what?”*

The estimate column is what the team lead sees on 14 October, priced at the rates configured in Kong before any invoice discount. The settled column is what finance charges back after close. Each estimate is scaled by its model's settlement factor, 0.94 overall, and the settled rows add up to $520 of AWS, $310 of Google Cloud, and $170 of Azure invoices. None of the per-user rows exist in the cloud bills.

Kong shows the estimate all month, by cost center and provider, and warns when a budget is on pace to run over. Your FinOps platform shows the settled number after close.

Kong knows who used which model shortly after each request. The invoices know what that usage really cost, but only after each provider closes the month. A monthly true-up joins the two by scaling Kong's estimates to the invoice totals, so you never have to push user identity into the clouds.

1. - Group the model-inference lines from the invoice and the records from Kong by **month, provider, model, and region**. Both sides already have these fields, so the clouds need no identity setup.
2. - For each group, divide the invoice amount by Kong's estimate to get the settlement factor. In October, for example, the organization's Bedrock Claude Sonnet invoice was $18,400 against a Kong estimate of $19,600, a factor of 0.94.
3. - Multiply each Kong record in the group by its factor. Each record then has an estimate and a settled cost.
4. - Put invoice lines with no Kong match on a central AI line that no team owns. That covers account-wide credits, traffic that didn't go through Kong, and fixed costs such as provisioned throughput, which stay there unless you choose a rule that splits them by usage share. The size of the line also shows how much AI traffic still bypasses the gateway.

• - Identity lives only in Kong. A request carries either an SSO (OIDC) identity, whose groups name the department, or a gateway key that belongs to a consumer group. The Kong attribution policy also turns an ai-cost-cost-center request header into a cost-center tag.

• - Kong uses your negotiated rates. Set your rates per million tokens (input_cost, output_cost, and the cache rates) on each model in Kong AI Gateway. Kong prices every request with them, which keeps the factors close to 1.0 and the mid-month numbers close to the final ones. If a factor drifts well away from 1.0, check that the rates in Kong still match your contract.
• - A model map ties each invoice line to Kong's provider and model names, and the key differs by cloud. On AWS it's ServiceName plus SkuMeter, and Anthropic models on Bedrock bill through AWS Marketplace under their own ServiceName, not Amazon Bedrock. On Azure it's a pattern over SkuMeter's abbreviated meter names. Google Cloud leaves SkuMeter empty, so there it's ChargeDescription or SkuId.
• - Map every token type to its model, including cache reads and writes, and drop region from the key wherever the two sides spell it differently.
• - Google Cloud keeps credits in x_Credits instead of Credit rows, so subtract them on the invoice side before you calculate any factors.
• - Nothing is counted twice. Kong's FOCUS 1.4 rows carry BilledCost 0, and in team reports the settled cost replaces the cloud's inference lines instead of adding to them.

To enable daily export of FOCUS data from Kong AI Gateway to your FinOps platform, configure an export schedule via the API:

Kong doesn't push files to your storage. You download each finished export through POST /v1/ai-cost-management/exports/{exportId}/downloads, which returns a URL that's valid for 5 minutes, so the pipeline starts with a small job that copies each file into your own bucket.

The real-time layer needs no export, since it's Kong's own usage views, budgets, and alerts. Exports are in beta in the public API (v0.3.0). On the cloud side, AWS FOCUS 1.2 billing data and Azure FOCUS 1.0 are generally available, while Azure FOCUS 1.2 and Google Cloud's FOCUS export to BigQuery are in preview.

To connect Kong with your FinOps platform, you bring together real-time telemetry from the gateway and monthly billing files from your cloud and model providers. The gateway tracks who used which model with request-level granularity, while cloud FOCUS exports and vendor billing feeds deliver the finalized costs after month close.

The following architecture illustrates how telemetry flows from cloud providers and Kong AI Gateway into your central FinOps platform.

Select the integration approach below that best aligns with your FinOps platform's ingestion capabilities and financial workflows.

1. - Route model traffic through Kong AI Gateway, with SSO identity or gateway keys mapped to cost centers, and set your negotiated rates on each model.
2. - Create the Kong export schedule shown above, plus a job that copies each file to your bucket.
3. - Connect the AWS, Azure, and Google Cloud billing data, and the OpenAI and Anthropic billing, to your FinOps platform.
4. - Build the model map from each provider's SKUs or usage types to Kong's provider and model names.
5. - Set up the true-up with the pattern your platform supports.
6. - Map cost centers to teams and publish the team views, with a dashboard that tracks the factors and the central AI line.
7. - Run it for one team for one month before you charge anyone back on settled numbers. You're ready when the settled total matches the invoice lines to the cent, the factors sit within a few percent of 1.0, and you can explain what's left on the central AI line.

If showback is all you need, skip the true-up and report Kong's estimates as they are. That works well when your Kong rates sit within a few percent of the invoice.

If you already tag cloud resources per tenant, you can carry a tenant ID into each cloud instead, through Bedrock inference profiles, Vertex AI request labels, or one Azure OpenAI resource per tenant, and then split each tenant's invoice lines by Kong's shares.

[_Kong AI Cost Management API_](https://developer.konghq.com/api/konnect/ai-cost-management/v1/)_Kong AI Cost Management API_  ·  [_Kong AI Cost Management_](https://konghq.com/products/ai-cost-management)_Kong AI Cost Management_  ·  [_FOCUS 1.4_](https://focus.finops.org/docs/specification/v1-4/)_FOCUS 1.4_  ·  [_AWS Data Exports for FOCUS_](https://aws.amazon.com/about-aws/whats-new/2025/11/aws-data-exports-focus-1-2-available/)_AWS Data Exports for FOCUS_

**What is AI FinOps?**

AI FinOps is the practice of applying cloud financial management principles to artificial intelligence and large language models (LLMs). It involves tracking, allocating, and optimizing AI spend in real-time. Because LLM costs fluctuate based on token usage rather than predictable server uptime, AI FinOps relies on specialized gateways (like Kong) to meter token usage, enforce budgets, and attribute costs back to specific business units or agents.

**How do I implement LLM chargeback without cloud tags?**

You can implement LLM chargeback without relying on native cloud resource tags by routing all model traffic through an AI Gateway. Kong AI Gateway attaches SSO (OIDC) identities, gateway keys, and cost-center tags to every request. This identity data lives exclusively in Kong. At the end of the month, a "true-up" process scales Kong's user-level usage estimates against the final cloud invoice, accurately charging back teams without ever pushing user identities into AWS, Azure, or Google Cloud.

**How do LLM budgeting solutions differ from traditional cloud budgeting?**

Traditional cloud budgets typically rely on billing data that lags by 24 to 48 hours, which is too slow for generative AI. A single runaway LLM agent can incur massive costs in minutes. Purpose-built LLM budgeting solutions operate in the request path. They meter tokens in real-time and can automatically enforce limits (e.g., returning an HTTP 429 error) the moment a dollar or token threshold is breached, actively preventing overnight cost spikes.

**How does the FOCUS billing spec apply to AI models?**

The FinOps Open Cost and Usage Specification (FOCUS) provides a standardized schema for billing data. For AI models, FOCUS normalizes complex, multi-provider billing (such as differing token metrics across OpenAI, Anthropic, Bedrock, and Azure) into a single, consistent format. Using the FOCUS 1.4 schema allows FinOps teams to process AI gateway exports and cloud invoices uniformly, making the monthly true-up and cross-model cost comparisons mathematically straightforward.

**How can I detect and stop an LLM cost spike overnight?**

To stop an overnight cost spike, you must move visibility from the invoice layer to the gateway layer. By routing traffic through Kong AI Gateway, every request is metered and priced instantly. You can set up the AI Rate Limiting Advanced policy to enforce hard token or dollar caps per consumer. If an automated coding agent or chatbot gets stuck in a loop overnight, the gateway will block the traffic immediately once the budget is hit, rather than surprising you weeks later on your cloud bill.

**How do I reconcile an OpenAI invoice with internal usage data?**

Reconciling an OpenAI invoice requires a monthly "true-up" algorithm. First, group your internal gateway usage records and the OpenAI invoice lines by month, provider, and model. Next, divide the actual invoice amount by your gateway's estimated cost to generate a "settlement factor" (e.g., 0.94). Finally, multiply your internal usage records by this factor. This adjusts your internal estimates to match the invoice exactly, while cleanly isolating any unmatched fixed costs or credits.