GreenOps & FOCUS: TokenOps' Environmental Twin
FOCUS: the FinOps standard extends to AI
The FinOps Open Cost and Usage Specification (FOCUS) reached v1.4 in June 2026 and now has a dedicated "FOCUS for AI" workstream to standardise token-based billing data across vendors. FOCUS 1.5 (in progress) explicitly scopes AI cost accountability improvements.
This matters because AI/token spend does not map cleanly to traditional cloud cost categories:
- Cost driver is tokens, not compute-hours.
- Same "resource" (a model call) has three price components (input, cached input, output).
- Reasoning tokens are output-priced but conceptually distinct.
- Fine-tuning, batch, and real-time pricing tiers coexist.
Enterprise signal: Walmart is publicly using FOCUS to unify multi-cloud and AI cost accountability — a leading indicator that Fortune 500 FinOps teams are extending existing cloud governance to token spend rather than building parallel systems.
GreenOps: same techniques, second scoreboard
The 2025–2026 research wave (ACL 2025, arXiv, Google Research) has produced a consistent finding: the same techniques that cut TokenOps dollars also cut carbon and water footprint.
- "Unveiling Environmental Impacts of LLM Serving" (ACL 2025) argues for standardising environmental impact per functional unit (useful response) — not per raw token — because context length and reasoning-token volume distort per-token comparisons.
- "How Hungry is AI?" (arXiv 2505.09598) benchmarks energy, water, and carbon across LLM inference — introducing water as a third accounting axis for ESG reporting.
- Google's "Measuring the Environmental Impact of Delivering AI at Google Scale" (arXiv 2508.15734, Patterson & Dean) is emerging as the de facto methodology enterprises cite for per-query energy/carbon estimation.
- CMU/Hugging Face "Energy Considerations of LLM Inference" (ACL 2025) ties efficiency techniques (caching, speculative decoding, quantisation) directly to measurable energy reduction.
The dual-reporting playbook
For each 2026 TokenOps optimisation, report both scoreboards:
| Optimisation | Dollar impact | Carbon impact (typical range) |
|---|---|---|
| Prompt caching (90% hit rate on 30% of tokens) | −25% $ | −25% CO₂e |
| Route 70% of traffic to Nano/Haiku tier | −40% $ | −35% CO₂e (smaller models = less energy) |
| Context pruning (halve avg context) | −30% $ | −25% CO₂e |
| Reasoning-tier governance (right-size effort) | −15% $ | −20% CO₂e (thinking tokens are energy-heavy) |
What to add to your 2026 reporting
- Cost per useful output, not cost per token.
- Energy per useful output (kWh), using Google/CMU methodology.
- Water per useful output — increasingly requested by ESG teams.
- Cached-token ratio — a leading indicator of both cost efficiency and carbon efficiency.
- Reasoning-token ratio — high reasoning ratio is a red flag on both scoreboards.
The convergence
GreenOps is not a parallel discipline; it is TokenOps reported against a second axis. Teams that mature their FinOps-for-AI practice inherit their GreenOps reporting nearly for free — the metrics, dashboards, and interventions are the same.