A typical AI cost conversation: finance sees one large invoice from a model provider, another from a GPU cloud, a growing line for AI features in SaaS tools, and a Kubernetes bill that now includes GPU nodes. Engineering can't say which product or customer drove any of it. Without that, AI unit economics are guesswork.
What FOCUS is
FOCUS, the FinOps Open Cost and Usage Specification, is the FinOps Foundation's open standard for billing data. The idea is simple: if every provider publishes cost and usage in the same columns, you can combine cloud, SaaS and AI spend in one dataset without writing a custom parser for each.
The standard has moved quickly. Version 1.3 added split cost allocation and a contract commitments dataset25. The FinOps Foundation reports that its steering committee ratified version 1.4 on 4 June 2026, adding new datasets and columns1, and coverage from FinOps X described additional data integrity tooling3.
Where AI fits, and where it doesn't yet
AI usage can be represented with existing columns: token counts as a consumed quantity and unit, the model and price point as a SKU, and AI services under a service category1. That is enough to put model API spend next to everything else.
It is not yet enough for detailed token economics. Reports around FinOps X 2026 described plans for version 1.5 to break down AI spend by token type and workload4. Until that lands and providers adopt it, input versus output tokens, cached tokens and per-feature attribution mostly have to come from your own telemetry.
| Cost source | Where the data comes from today |
|---|---|
| Hyperscaler AI services | Provider billing exports, increasingly in FOCUS format |
| Model API providers | Invoices and usage APIs; detail varies by provider |
| Self-hosted models on Kubernetes | Cluster cost allocation, with GPU attribution per workload5 |
| AI features in SaaS tools | Seat and add-on licences, often outside any usage feed |
| Per-feature and per-customer usage | Your own gateway and tracing telemetry |
The architecture I use
- A model gateway for every call, which records model, input and output tokens, latency and the calling team, product, feature and, where relevant, customer.
- Tracing with GenAI attributes, as in LLM observability with OpenTelemetry, so cost can be tied to user journeys rather than just API keys.
- FOCUS as the finance format. Normalise provider exports to FOCUS, and publish internal usage in compatible columns so finance sees one dataset.
- GPU allocation for self-hosted models, using the cluster techniques in Kubernetes FinOps and DRA for GPUs.
- Reconciliation each month between gateway totals and invoices. Gaps reveal untracked keys and shadow usage.
The single most useful number is cost per successful outcome: per resolved ticket, per document processed, per qualified answer. Tokens are an input; outcomes are what the business buys.
Decisions this unlocks
- Model routing. Send simple requests to cheaper models once you can see which features use expensive ones.
- Build versus buy. Compare API spend with the cost of self-hosting, as in build vs buy for AI, using real usage instead of estimates.
- Pricing. Set customer pricing for AI features with actual unit costs, not hopes.
How I can help
I set up AI cost visibility and unit economics, from gateways and tagging to FOCUS-based reporting finance can use. As a fractional CTO I can own the cost roadmap alongside the technology one. Or analyse your challenge and bring the result to a call.
- 1FOCUS SpecificationFinOps Foundation
- 2FOCUS Specification v1.3 (PDF)FinOps Foundation · Dec 2025
- 3FinOps Foundation extends specification to rein in cloud costs in the AI eraTechstrong IT · 2026
- 4FOCUS specification and AI cost accountability at FinOps XSiliconANGLE · 8 Jun 2026
- 5