Imported from faionfaion/faion-network (skills/faion/knowledge/infra/finops-ai-ml-costs/AGENTS.md). Install upstream with npx skills add faionfaion/faion-network --skill finops-ai-ml-costs. Copyright stays with the author.

FinOps for AI / ML Workloads

Summary

One-sentence: Generates an AI/ML cost-control config (GPU spot strategy with checkpointing + inference caching + quantization + batch policy + per-run/per-request attribution) for training and inference.

One-paragraph: Generates an AI/ML cost-control config (GPU spot strategy with checkpointing + inference caching + quantization + batch policy + per-run/per-request attribution) for training and inference. The methodology pins the artefact shape, ties every conclusion to a rule, and routes the operator via a decision tree that always terminates either on an applicable rule or on skip-this-methodology. Apply when preconditions hold; skip via the tree otherwise.

Ефективно для:

• GPU training pipelines з runtime ≥1h і checkpointable jobs.
• Inference fleet з cacheable prompts (RAG, classification, embeddings).
• Multi-tenant ML platforms де per-team chargeback потрібен.
• Pre-deployment review нового model: cost-per-prediction must clear ROI gate.

Applies If (ALL must hold)

• GPU spend ≥$5k/month OR cloud LLM inference spend ≥$2k/month.
• Training jobs support checkpoint/resume.
• Per-team or per-product attribution required by finance.
• Inference traffic has measurable repeat-input rate (≥10%) to justify caching.

Skip If (ANY kills it)

• GPU spend <$2k/month — vendor pricing optimization gives bigger ROI than spot.
• Training jobs cannot checkpoint (research scripts in flux) — spot is unsafe.
• Latency-critical inference (<100ms P99) — caching may add coherence drift unacceptably.

Prerequisites

Assumes Loaded

Content (load on demand)

Task Routing

Templates

Scripts

Related

• Parent: pro/infra/cicd-engineer/AGENTS.md
• [[finops-framework]]
• [[gitops-core-principles]]

Decision tree

See content/06-decision-tree.xml. The tree starts from a concrete observable signal and routes each branch to a <conclusion ref="rule-id"> resolved against content/01-core-rules.xml. Use it whenever you are unsure whether this methodology applies — the tree always terminates either on an applicable rule or on skip-this-methodology.