FinOps Foundation

On 17 February 2026 the FinOps Foundation published FinOps for AI Overview, a guide co-written by a working group drawn from Google, AWS, MetLife, Wells Fargo, Roche, Accenture, KPMG and EY, released under CC BY 4.0. It treats the token as the fundamental unit of consumption and maps the 14 capabilities of the FinOps Framework onto what stays common with cloud and what changes for AI. Allocation is the capability most disturbed: tracing multi-agent workloads has no standard framework.

The KPIs come with formulas and worked figures. Cost Per Token, Cost Per Inference, Training Cost Efficiency and ROI sit alongside an LLM Model Choice Quality Score Alignment that compares the MMLU score a task actually requires against the MMLU of the model running it, so a sentiment analysis needing 54 is visibly oversized on GPT-4. Maturity follows Crawl, Walk, Run; eight pricing models are catalogued; showback comes before chargeback. A Certified FinOps for AI certification accompanies the guide.

The 30× to 200× cost variance the FinOps Foundation observes between an unoptimized and a well-optimized deployment is the figure other vendors reuse: Finout builds its November 2025 CPO guide around it. In June 2026 the Linux Foundation announced the Tokenomics Foundation, operated in close partnership with the FinOps Foundation and committed to extending its FOCUS specification, described as carrying the discipline of variable technology spend into the era of token-based AI. The February guide still does not treat AI agents as a category of their own.