FinOps & GreenOps
A cloud bill nobody can attach to a workload is not a cost problem, it is a readability problem. Once every euro has an owner, reduction stops being a quarterly debate and becomes a decision the teams make themselves.
Nobody reduces a spend they cannot attach to anyone.
Formal cloud cost programmes went from thirty-nine to seventy-two per cent of organisations in a single year, and ninety-eight per cent of FinOps practitioners now manage AI spend, against thirty-one per cent the year before. The subject has moved from the finance department to the engineering teams, which is where the decisions actually are.
What we do is put the spend where those decisions are made. Tagging that refuses a resource without an owner, a budget per workload rather than a global envelope, a cost per request on AI workloads, and a footprint tracked at the same pace as the invoice. The arbitration itself stays with the teams: they are the only ones who know whether that test environment will be needed again in October.
What we do
Four workstreams, in the order that makes the next one possible.
Make the spend readable
Tagging enforced at creation, spend attached to a workload and to an owner, showback before chargeback. Re-invoicing a team for something it cannot act on produces resentment, not savings.
a resource without an owner is refused, not billed
- Tags enforced at creation, not audited after
- Spend attached to a workload and a team
- Showback first, chargeback where the lever exists
Reduce without degrading
Sizing on real demand, reserved baseline set on the trough, retention degraded by age, environments with a declared end date. The savings are in the policies, rarely in the machine sizes.
each reduction goes back to the team that owns it
- Reserved baseline on the trough, elasticity on the peak
- Retention degraded by age rather than uniform
- An end date declared at environment creation
Control the cost of AI
Cost per completed request rather than per token, routing by difficulty so the largest model is used only when the smaller one fails, and a cutoff that queues rather than fails.
no scaling without two weeks of measured real usage
- Cost per request, per model and per use case
- Routing by difficulty, context cached
- A budget threshold that queues instead of failing
GreenOps, measured not claimed
Footprint tracked at the same pace as the bill: requests, transferred weight, resource use per journey. Eco-design gives most of its result on the same levers that lower the invoice, which is what makes it hold.
measured on the journey, not on a yearly report
- Footprint per journey, measured per release
- Eco-design on requests and transferred weight
- The lines where cost and footprint fall together
What you get
One project runs through the four deliverables below: bringing a platform's cost under control. Each line states what is actually handed over, in the order it is handed over.
The tagging policy, enforced
Four mandatory tags, refused at creation rather than audited afterwards, and a catch-up campaign by domain on what already runs. Without this, every later figure is a discussion rather than a fact.
A budget per workload
A threshold per workload rather than a global envelope, an alert at eighty per cent of the trajectory rather than at a hundred per cent of the amount, and drift caught in days instead of on the monthly statement.
AI workloads at cost per request
Cost per completed request, per model and per use case, with routing by difficulty and a cutoff that queues. A successful pilot is also a bill that takes off.
A monthly arbitration, not a cut
Three lines concentrate most of the gap, and they go back to the teams that own them with the amount next to them. What the execution finds is not a decision, it is what a decision needs.
How we deliver
Make it readable
depending on the number of accounts, teams and the state of tagging
- Complete tagging, team and workload
- Spend attached to an owner
- The drifting lines, ranked by amount
Arbitrate
depending on the number of workloads and the contractual room to move
- A threshold per workload, agreed with the team
- First reductions decided, not imposed
- Retention and environments reworked
Extend
depending on the teams to bring on board and the share of AI workloads
- AI workloads tracked at cost per request
- Eco-design measured on the journeys
- Cost in the team review, not in a committee
Hold
service commitment defined with you
- Drift caught in days, not on the statement
- Thresholds revised as usage changes
- Footprint tracked at the same pace as the bill
What it changes in your business
The line that drifts is never the same. Continuous ingestion in energy, seasonal peaks in retail, duplicated environments in finance, request volume on AI workloads. A page that talked about cloud cost in general would say nothing to anyone.
Ingestion that never stops
Readings arrive continuously, around the clock, and the bill follows the same curve. What costs is not the compute but the retention: time series kept at the second for thirteen months when the minute is enough after a week. The saving is in the retention policy, not in machine sizing.
- Retention degraded by age, not uniform
- Storage cost attached to its domain
- Compute moved to the data, not the reverse
Where cost management actually stands
Bills that go down while the platform grows
A cloud bill that goes down while the platform grows
−24% cloud bill · 100% Terraform IaC
A group of 7,000 people whose cloud spend nobody could attach to a team, on an estate deployed by hand where every environment had drifted from the next.
A cloud-native factory on managed Kubernetes with the whole infrastructure described in Terraform. Resources tagged from the start, so spend attaches to a workload and an owner rather than to an invoice line.
A bill that falls while usage multiplies
−22% cloud costs · ×4 user capacity
A user base growing towards 200,000 active customers, on an AWS estate whose invoice was following the same curve, with no way to say which part of the growth was paying for itself.
Continuous optimisation of AWS consumption, rightsizing, autoscaling and spot capacity, paired with carbon tracking per business journey. The bill falls by 22% while the platform absorbs the growth.
Insights & Perspectives
Cloud-native infrastructure: the foundation of modern enterprises
Kubernetes, platform engineering, GitOps, supply chain security and FinOps: why cloud-native infrastructure has become the foundation of modern enterprises.
Green IT and observability: making the transition work
Sustainability metrics and observability for green IT: PUE, data throughput, resource utilisation, an observability stack and five steps to a sustainable organisation.
Is Kubernetes right for you?
Kubernetes orchestrates containers at scale but brings complexity, cost and a steep learning curve. Architecture, benefits, limits and a method to decide in 2026.
Attach the spend, then decide
Tagging enforced at creation, a budget per workload, AI workloads measured per request, and an arbitration that stays with the teams.
Frequently asked questions
Because a spend nobody can attach to a team cannot be arbitrated. It gets discussed every quarter without a conclusion. Tags enforced at creation turn the invoice into a set of lines each with an owner, which is what makes a reduction decidable.
Showback shows a team what it spends, chargeback bills it. We start with showback everywhere, and only move to chargeback where the team actually holds the lever. Billing someone for something they cannot change produces resentment, not savings.
From policies far more often than from machine sizes. Retention degraded by age, a reserved baseline set on the trough rather than on the peak, and environments with a declared end date usually weigh more than any rightsizing exercise.
By measuring cost per completed request rather than per month, model by model and use case by use case, then routing by difficulty so the largest model is only called when a smaller one fails. One organisation in five overshoots its AI forecast by more than half.
No. Execution produces the lines, the amount and the owner, and the arbitration goes back to the business. Only the team knows whether a test environment will be needed again in October, and cutting it without asking is how a programme loses its credibility.
Often, not always. Fewer requests, less transferred weight and better resource use lower both. Moving a workload to a greener region lowers the footprint without lowering the bill. We report the two curves separately, and name the lines where they meet.
Making it readable takes 2 to 6 weeks depending on the number of accounts, teams and the state of tagging, and produces complete tagging, spend attached to an owner and the drifting lines ranked by amount. The first arbitrated reductions follow in 4 to 10 weeks.