Ten questions, then a range in euros.
Answer what you know. “I don't know” is a fine answer: the range gets wider, not wrong. You see where savings usually sit on a bill like yours, level by level, and where to look first.
The ten questions
The diagnostic runs in your browser and needs JavaScript. Here are its questions; how the range is worked out, with every figure it uses, is just below.
- How much is the cloud bill each month?
- What weighs most on it?
- How much of it is non-production?
- Is non-production turned off at night and at weekends?
- How much of the steady compute is under commitments?
- When were machines and databases last resized to real usage?
- Does anything run on ARM?
- Does old data move to cheaper storage tiers?
- NAT gateways and data transfer: how big a line?
- Do you pay for LLM APIs or GPUs?
How much is the cloud bill each month?
All clouds together, after discounts, excluding tax. A rough figure is enough.
What weighs most on it?
A guess is fine.
How much of it is non-production?
Development, test, staging, demos.
Is non-production turned off at night and at weekends?
How much of the steady compute is under commitments?
Savings Plans, reserved instances, committed use discounts.
When were machines and databases last resized to real usage?
Does anything run on ARM?
Graviton on AWS, Axion on Google Cloud, Cobalt on Azure.
Does old data move to cheaper storage tiers?
Lifecycle rules on buckets, archive tiers for backups.
NAT gateways and data transfer: how big a line?
Do you pay for LLM APIs or GPUs?
Your answers
Change any of them: the range follows.
How the range is worked out Sources and assumptions
Split of the bill assumed from your answer on the biggest line: compute 45%, databases 20%, storage 12%, network 8%, AI 5%, the rest 10%.
Each lever takes a share of the spend it acts on, one after the other in the order of the method's levels, so nothing is counted twice. Each answer sets a low and a high share; “I don't know” keeps the widest one, so the range gets wider rather than wrong.
Levels 4, 6 and 7 work on contracts, code and decisions. They need your bill and your teams, so they stay out of the range.
For comparison: about 35% of a cloud bill usually comes back without touching the architecture.
Rates from published sources
- 76.2% Share of the week a development or test environment can be off: used 40 hours of 168. AWS Well-Architected, cost optimization design principles, quoted in the Amazon RDS article
- 24.2% One-year No Upfront Compute Savings Plan against On-Demand, m7g.large in Paris. AWS Price List, eu-west-3, fetched 2026-10-14, in the AWS EC2 article
- 34% One-year No Upfront EC2 Instance Savings Plan against On-Demand, m7g.large in Paris. AWS Price List, eu-west-3, fetched 2026-10-14, in the AWS EC2 article
- 19% Graviton against Intel, On-Demand: m7g.large against m7i.large in Paris. AWS Price List, eu-west-3, fetched 2026-10-14, in the AWS EC2 article
- 45.4% S3 Standard-IA against S3 Standard, per GB-month in Paris. AWS Price List, eu-west-3, fetched 2026-10-15, in the Amazon S3 article
- 79.2% S3 Glacier Instant Retrieval against S3 Standard, per GB-month in Paris. AWS Price List, eu-west-3, fetched 2026-10-15, in the Amazon S3 article
- 50% Batch inference against on-demand: "at a 50% lower price". Amazon Bedrock pricing page, quoted in the Amazon Bedrock article
- 75% Cache read input tokens against on-demand input tokens: "75% less" (Nova 2 Lite). Amazon Bedrock pricing page, quoted in the Amazon Bedrock article
- 20% My fee: a share of the savings measured on the bill. The FinOps offer
- 35% The share of a cloud bill that usually comes back without touching the architecture. The FinOps offer
The model's assumptions
- Split of the bill, by its biggest line
- Compute: virtual machines, Kubernetes: compute 55%, databases 17%, storage 10%, network 7%, AI 3%, the rest 8%. Databases: compute 35%, databases 35%, storage 12%, network 7%, AI 3%, the rest 8%. Storage: compute 35%, databases 17%, storage 30%, network 7%, AI 3%, the rest 8%. Network and data transfer: compute 38%, databases 17%, storage 12%, network 22%, AI 3%, the rest 8%. AI: LLM APIs, GPUs: compute 30%, databases 14%, storage 8%, network 5%, AI 35%, the rest 8%. I don't know: compute 45%, databases 20%, storage 12%, network 8%, AI 5%, the rest 10%.
- Non-production's share of compute and databases
- Under 10%: 5–10% · 10 to 25%: 10–25% · 25 to 40%: 25–40% · Over 40%: 40–50% · I don't know: 10–35%
- Share of non-production already off at night and at weekends
- Yes, all of it: 100% · Some of it: 40–60% · No, it runs all the time: 0% · I don't know: 0–60%
- Hours a week non-production stays up, low case
- 60 hours of 168 (twelve hours a day, five days); the high case is the 40 hours above
- Share of a stopped database's cost that stops with it
- 50%: storage and backups keep billing
- Orphaned volumes, old snapshots and idle IPs
- 1–3% of compute and storage
- Rightsizing, a share of compute and databases
- In the last 3 months: 2–5% · Earlier this year: 5–12% · Over a year ago: 10–20% · Never: 15–30% · I don't know: 5–25%
- Share of compute that can still move to ARM
- Most of it: 0–10% · Some services: 15–40% · No: 40–75% · I don't know: 15–75%
- Share of the steady base still to put under a commitment
- None: 75–90% · Under a third: 40–60% · About half: 20–35% · Most of it: 0–12% · I don't know: 12–75%
- Commitments on databases
- the EC2 Savings Plan rates above, applied to databases too
- Share of stored data that has cooled and not moved to a cheaper tier
- Yes, everywhere: 0–6% · On some buckets: 20–25% · No: 40–50% · I don't know: 10–40%
- Share of the network line that endpoints and placement remove
- A big line: 20–40% · A small line: 5–15% · Barely there: 0–5% · I don't know: 5–30%
- Share of AI spend that caching, batch or a smaller model can take
- No: 0% · Yes, without caching or model routing: 60–75% · Yes, already optimised: 10–20% · I don't know: 20–55%
- Ceiling of the high end
- 60% of the bill
The audit puts your own figures on it.
Free, 48 hours, from your billing exports and with your negotiated discounts. I'm paid 20% of the savings measured on your bill, for 12 months, and nothing if the bill does not go down.