The AI bill grows faster than usage.
The right model for each request, and a cache that is actually read. Routing and prompt caching, tested on your own traffic.
Levers · routing · prompt caching · batchI cut AWS, Azure and Google Cloud bills without slowing production down. The audit is free and takes 48 hours. My fee is 20% of what the bill actually loses.
A bill you can explain line by line, and a fee that exists only if it drops.
Savings that ship as engineering changes, checked against performance before production.
Field notesAbout 35% of a cloud bill usually comes back without touching the architecture.In 48 hours there are always three levers, before Spot and reservations.An AI bill that explodes is mostly the wrong model per request, and a cache that writes without reading.The big mistake with reservations is not using all of what you reserved.Leaving the cloud only pays for steady, recurring loads.
Six situations behind a cloud bill that grows faster than it should. Each has its levers, and none of them needs a slower system.
The right model for each request, and a cache that is actually read. Routing and prompt caching, tested on your own traffic.
Levers · routing · prompt caching · batchRequests, autoscaling and node pools measured on real usage, and bare metal through Kubernetes where the load is steady.
Levers · rightsizing · autoscaling · bare metalProfile first, then rewrite or move the slow part. At EDF, a 7 h 30 run came down to under 30 seconds.
Levers · profiling · parallel runs · ARMSavings plans and reservations sized to what actually runs, so all of what you reserve gets used.
Levers · coverage · utilisation · termTransfer and NAT priced before anything moves. The cheaper answer often stays put, with storage tiers and free routes.
Levers · NAT · egress · tieringCosts mapped to products and teams, with a unit cost finance can follow month after month.
Levers · tagging · allocation · unit costNot sure which one is yours? The bill diagnostic tells you in three minutes.
Run the bill diagnostic LiveEvery saving comes as an engineering change: what to change, the effort, the risk for production. Your team approves each one.
I write Rust and design Kubernetes platforms and data pipelines. When a change needs code, I can write it with your engineers.
Production is where I come from: a real-time IoT platform for over a million sensors, energy simulation at EDF, platform engineering at Pierre & Vacances Center Parcs.
of the measured saving, for 12 months. If the bill does not go down, you owe nothing.
€1.3M saved on client bills in total. No amount is attributed to a single client.

EnergyA real-time IoT platform, and a simulation engine moved from C# to Rust.

LeisurePlatform engineering: a developer plugin and a bug telemetry platform.
RailValidation on high-speed and intercity trains, in Europe, Egypt and the United States.

RetailThe run and evolution of an on-premises Hadoop data lab.
Cloud architect and FinOps engineer, self-taught, writing code for clients since 2019. I read every bill myself, and I treat cost as part of the design rather than a surprise after it.
Seven levels to work on, four steps to get there, and one optional step when the audit finds bigger work. Nothing changes in production without your team's approval.
Seven levels, from cutting what serves no purpose to deciding where the money goes. Tools and checklists cover the first three. The savings that last sit further down, in architecture, code and steering.
| Workstream | Levers | Effort | Timeline |
|---|---|---|---|
| 1HygieneCut what serves no purpose | Orphaned volumes and IPs, old snapshots, non-production off at night and at weekends, tags. | Low | Hours to days |
| 2RightsizingFit to real usage | Oversized instances and databases, gp2 to gp3 disks, a move to ARM (Graviton). | Medium | Days to weeks |
| 3CommitmentsPay less for what remains | Savings Plans, reserved instances, Spot for workloads that can take an interruption. | Low | Days |
| 4NegotiationCloud and vendor contracts | Multi-year commitments (EDP, MACC), private pricing, migration credits, licences and SaaS contracts. | Medium | Weeks to months |
| 5ArchitectureChange what generates the bill | Managed or serverless services, autoscaling, caching, outbound traffic, storage lifecycle. | High | Weeks to months |
| 6Code and runtimeRewrite what runs the most | Profile the hot paths, change language where it pays. A C# batch moved to Rust went from 7 h 30 to under 30 s. | High | Months |
| 7CTO steeringDecide where the money goes | Cost per customer or per transaction, the cloud's share of revenue, build or buy, repatriation where it is justified. | High | Quarters |
| Each level works on what the levels above it left. | |||
Five steps, from the first call to the bigger work. At step 3, the example bill goes through the seven levels above.
Book a call and bring the last bills, or just their totals and your main services.
I tell you whether the offer fits, before any paperwork. If it does, we sign an NDA.
A straight answer on fit, and an NDA before any data changes hands.
Read-only access to billing is enough. No access to production at this stage.
Share the billing exports, read-only.
I read the bill line by line across the seven levels, and price each lever with its effort, its risk and what it does to production.
The lever table for engineering and a one-page summary for finance.
We walk through both with your CTO and your CFO on the same call, and your team picks what to apply.
| Level | Lever | Saving a month | Effort | Risk | Production |
|---|---|---|---|---|---|
| 1 | Schedule test environments | €7,400 | 1 day | Low | No production change |
| 1 | Delete old snapshots and orphaned volumes | €4,900 | 1 day | Low | No production change |
| 2 | Rightsize the main compute group on ARM | €12,600 | 1 week | Medium | Tested on staging first |
| 3 | Three-year commitments on the stable compute | €18,900 | 2 days | Low | Three-year commitment, stable floor only |
| 4 | Negotiate a multi-year discount (EDP) | €9,950 | 1 week | Medium | No production change |
| 5 | Prompt caching and model routing | €6,500 | 2 weeks | Medium | Quality and latency measured |
| 5 | Gateway endpoint for S3, not the NAT | €3,000 | 1 day | Low | Tested on staging first |
| 6 | Incremental export, smaller database | €6,800 | 1 week | Medium | Tested on staging first |
| 7 | Retire the reporting stack nobody opens | €4,200 | 2 days | Low | Data archived first |
| Total | €74,250 | 32 days | 55.0% of the bill |
Pick the levers to apply. Your engineers review and deploy each change.
I write or drive each change with your teams: the pull request or the configuration, its test, and a way back.
A lower bill without a slower system. The fee starts only once a saving shows on the bill.
In this example, the €135,000 monthly bill goes down to €60,750 once the seven levels are applied: €74,250 a month saved, and a fee of €14,850 a month.
Keep scrolling: each level takes its share off the bill.
Pick a level to see the bill after it.
Keep shipping. Growth in usage is taken into account, not held against you.
Each month I measure the saving against the baseline, at comparable usage, and send the statement.
A monthly statement you can check line by line. After 12 months, every euro saved is yours.
The baseline comes from recent bills. Discounts and commitments you already had stay in it, so the fee only covers what changed.
The usage measure is agreed at the start: requests, users or data processed.
Decide whether the bigger work is worth it, with the figures in hand.
When the audit finds a redesign or a migration worth doing, I scope it as a separate project: goal, cost, expected saving and plan.
A fixed scope you can accept or not. The levers already found stay paid on results.
Levels 5 to 7, when a lever needs more than a setting: a batch rewritten and run in parallel, a platform moved to bare metal through Kubernetes, an AI feature rebuilt around caching and routing.
Exercises built on what each company published, priced at AWS list prices. Not client work: I have no relationship with these companies. The percentage is the base scenario's and covers every lever of the exercise; the levers named here are the level 5 ones.
An architect on your team, up to 20 days a month, in Paris or remote. Daily rate on request.
Every component priced before it is built. The cheaper option is kept when it does the job, and the monthly bill is estimated while the design is still on paper.
The architecture written as specifications your engineers, or their coding agents, can build from: decisions and their reasons, constraints, interfaces and acceptance criteria, before the first line of code.
Twenty articles so far, one cloud service at a time, each with its prices and its levers.
Worked on your own cases, in small groups or one to one.
Decisions, applying them, and the budget of a tech team, worked on your real cases.
10 hours · €2,400 · first hour freeCloud services, what they cost and who builds them, explained to the people who sign the budget.
Duration and price to setRead the bill, find the levers, and ship the change without hurting production.
Half a day to two days · from €1,500Moving the hot path from C# or Python to Rust, the way the EDF engine went from 7 h 30 to under 30 seconds.
Duration and price to setNot for the audit: it only needs billing exports, shared after an NDA. Access comes later, only for the changes your team approves, under your own rules.
Then you owe nothing. The fee is 20% of the measured saving, for 12 months, and only once the saving shows on the bill.
Against a baseline set from recent bills, at comparable usage. Discounts and commitments you already have are part of that baseline.
Performance is part of every review. A cheaper option that hurts latency or reliability is not counted as a saving.
AWS, Google Cloud and Azure, separately or combined, with Kubernetes and data platforms on top.
Yes, from January 2027: FinOps or architecture by the day, up to 20 days a month, in Paris or remote.
Free audit · NDA first · nothing changes without your approval