Insights · FinOps for AI, Azure, AWS & Google Cloud — from the people who build the checks
Deep, practical guides on finding and eliminating waste across AI workloads, Azure, AWS and Google Cloud — written by the team behind the CloudFinOpsKit tools' cost checks. Real meter behaviour, real CLI commands, real traps we've hit so you don't have to.
Token-denominated spend behaves differently from everything else on your bill. Start with the framework.
The hub page. The four things that genuinely make AI spend different — a virtual billing currency, tokens that cannot be tagged, cost coupled to quality, and spenders outside engineering governance — mapped onto the FinOps Framework domains, with a maturity ladder, a 90-day plan and a curated library of the primary sources.
Your bill is denominated in tokens; you write in words; the exchange rate is not fixed. The four multipliers that inflate it — language, serialisation format, identifiers, and context you never wrote — and why shortening your text can make it cost more.
You can't manage what you don't meter. The economics half of tokenomics — meter token usage, price the true blended cost per query, attribute AI spend to an owner, and govern it as agents multiply the bill.
Tokens are the atomic unit of AI cost. Measure tokenomics, then pull four levers — right-size & route the model, trim context, cap output, cache repeats — to cut LLM spend 60-80% without losing quality.
A cheaper model that fails twice costs more than an expensive one that works first time. How to define an evaluable success bar, instrument cost across retries and fallbacks, and turn model selection into a total-cost decision instead of a price comparison.
Agents replay their whole history on every step, so cost grows with the square of the step count — doubling the steps roughly quadruples the bill. The failure modes that produce runaway spend, and the layered controls that bound it without crippling the agent.
Why token spend breaks classic anomaly detection, and the four detectors that work: tokens/day spikes, output:input ratio bloat (output costs 3–5× input), blended $/1K-token rate shifts that expose a model-mix change, and idle provisioned throughput.
AI cost is the #1 FinOps focus of 2026. Why it breaks the old playbook, the levers that control it (prompt caching, right model, PTU, output caps, zombies), and how to attribute token spend.
These four options don’t just cost different amounts — they cost in different shapes. One is nearly all marginal, one is fixed. How to find your break-even, and the costs each option hides (including the GPU rate that isn’t the cost).
Per-model token visibility from CloudWatch, then five levers — prompt caching, output caps, model right-sizing, context trimming, and the right Provisioned Throughput call — to cut Bedrock spend without losing quality.
Get from the Vertex service total to the four things that move it: idle endpoints billing for nothing, output and context bloat, batch prediction for anything non-interactive, and when Provisioned Throughput actually pays.
The operating model, unit economics, governance and standards — the parts that are not specific to one provider.
Database savings plans on AWS and Azure, the end of Azure reservation exchanges in February 2027, Google’s switch from CUD credits to discounted rates, and FOCUS commitment data — and which persona owns each part of the return.
Executive Strategy Alignment, Scopes, FOCUS 1.4 and the Tokenomics Foundation — the five billing models FinOps now covers, and how every persona feeds one ROI number per Scope.
A CCoE isn’t six specialists advocating for six pillars — it’s the one place they’re allowed to argue. What it owns across AI, cost, security, reliability, operations and performance, the failure mode when nobody owns it, and the single artifact that proves ownership is real.
The preventive guardrails, detective review cadence and accountability loop that turn a one-off cleanup into a durable FinOps practice — plus a 90-day rollout.
The frontier of FinOps maturity — tie spend to a business outcome. Pick the right unit, compute cost per customer or transaction, and turn the metric into decisions.
Showback vs chargeback, the three ways to allocate Azure spend, handling shared costs nobody owns, and producing a per-team statement finance will actually act on.
Why every major cloud now emits FOCUS natively, the core columns, how to turn it on in Azure Cost Management, and why it makes your cost data AI-ready.
A sudden spike, investigated in order: confirm it's real (unfinalized days, commitment purchases, expired credits), isolate the service and the day it started, drill to the resource, match it to the usual suspects — data-transfer blow-outs, scale-outs, snapshot sprawl, leaked keys — and stop the bleed. With the exact Cost Explorer and Cost Management steps.
Azure's anomaly alerts and AWS's Cost Anomaly Detection compared — scopes, tunability, delivery, blind spots — and one auditable deterministic methodology (same-weekday spikes, step changes, Theil–Sen creep, new/vanished spend) that gives both clouds a single definition of "anomaly".
Close the gap between what your pods request and what they use: right-size requests, bin-pack with autoscaling/Karpenter, run Spot, scale non-prod to zero, and cover the baseline with commitments.
Where the money actually is on Azure, and the exact portal or CLI action for each.
A prioritized, do-it-in-order playbook: kill zero-risk waste first, turn off non-prod out of hours, rightsize, apply Hybrid Benefit, layer reservations & savings plans, tame Log Analytics, and make it a monthly habit — each step with the exact Azure action.
The complete working checklist — compute, storage, network, databases, commitments, licensing, monitoring and AI workloads — grouped the way a real cost review runs.
Built-in Cost Management anomaly alerts, the daily and month-over-month checks that matter, a threshold you can explain, and the usual suspects behind a spike.
Discount depth vs flexibility, use-it-or-lose-it utilization, exchange rules, the Dev/Test subscription myth, and the layering strategy mature teams use.
Who qualifies, who doesn't (Dev/Test subscriptions, Windows client VMs, SQL Express/serverless), the 8-core minimum, and the exact CLI to find every VM not using it.
The five tags that drive allocation, enforcing them with Azure Policy (require & inherit), backfilling an untagged estate, and tracking coverage as a KPI.
Portal, Resource Graph and CLI methods — plus the false positives that burn people: Site Recovery replica disks, backup snapshots, and disks on stopped VMs.
Run-rate vs trend vs driver-based methods, the built-in Cost Management forecast, tracking budget variance to ±10%, and the commitment and anomaly factors that throw forecasts off.
Azure Cost Management, Advisor, Microsoft's FinOps toolkit, CloudFinOpsKit, and the enterprise platforms — what each actually does, what it costs, and when you've outgrown it.
The same discipline on AWS — commitments, orphaned resources, anomaly detection and the tooling landscape.
The working checklist we run on AWS — kill waste (EBS, EIPs, idle NAT/LBs), right-size with Compute Optimizer, commit smart with Savings Plans, tier S3, govern with Tag Policies, and optimize Bedrock spend. Mapped to the Well-Architected Cost Optimization Pillar.
Set up free, ML-based anomaly monitors, pick a threshold you can explain, read the root-cause hints, and pair it with budgets + a monthly review to catch slow creep.
Compute SP vs EC2 Instance SP vs RIs — coverage, discount and the flexibility trade-off, how Graviton changes the math, and a coverage strategy that bounds stranded-commitment risk.
Find available EBS volumes with the Console, CLI and Cost Explorer — and dodge the traps: snapshots backing AMIs, volumes on stopped instances, idle Elastic IPs, and DR snapshots. With a safe, snapshot-first workflow.
Ten KPIs grouped by measure, optimise, govern and prove-the-return — each with a formula, a working band, and the part other lists skip: exactly how it can mislead you. Including why commitment coverage quoted without utilisation can be describing waste, and why falling total spend is the most commonly misread number in FinOps.
A decision framework rather than a vendor ranking: the five categories of FinOps tool and what each one genuinely cannot do, a table mapping the symptom you actually have to the category that addresses it, and the two questions no demo volunteers — how you get your data out, and how much spend lands in "unallocated".
An honest comparison: the free AWS-native tools (Cost Explorer, Compute Optimizer, Cost Optimization Hub), commitment automation, CloudFinOpsKit, and enterprise platforms — what each does, what it costs, and who it's for.
GCP's discount model is genuinely different — sustained use applies automatically, which changes the maths.
A prioritized, do-it-in-order playbook for GCP: kill idle resources first (Active Assist recommenders), turn off non-prod out of hours, right-size, layer CUDs and SUDs, tier Cloud Storage with Autoclass, cut Standard-tier egress, optimize GKE, and watch Vertex AI token spend — each with the exact Google Cloud action.
How GCP's two rate levers work and stack — automatic Sustained Use Discounts and opt-in Committed Use Discounts — resource-based vs spend-based (flexible) commitments, and a coverage strategy that covers the steady floor with CUDs while SUDs handle the variable top, without stranding a 3-year commitment.
GCP's native cost anomaly views and Cloud Billing budget alerts (with Pub/Sub), where they fall short — budgets alert on totals not causes, and gradual creep hides — and an auditable deterministic methodology (same-weekday spikes, step changes, Theil–Sen creep) run on the BigQuery billing export.
An honest comparison across five tiers: the free native tools (Cloud Billing, FinOps Hub, the Recommender API and the BigQuery billing export that everything reconciles against), GKE cost tooling, commitment automation, CloudFinOpsKit, and enterprise platforms — what each does, what it costs, and who it's for.
