The Baseline
Before the Model
32% lower cloud spend on one Azure landing zone. No model in the path.
A client asked which AI FinOps tool to buy. The answer was none yet.
Ne Plus Ultra Global Solutions | Engineering & Architecture
Which AI FinOps tool should we buy?
A client asked which AI FinOps tool to buy to fix cloud sprawl on an Azure landing zone.
None yet.
A model multiplies the architecture underneath it. On an estate that still holds idle nodes, repeats security scans, and admits portal writes, the multiplier arrives before the measurement. This paper records the baseline that was taken first, on one landing zone, with the model left out of the path.
Three planes, read before any license.
The audit covered node capacity, pipeline duration, and state ownership.
- ▪AKS node pools held a static 16-node minimum through off-hours.
- ▪Builds ran 40 minutes, with duplicate SAST on unmerged pull requests.
- ▪State. 18% of assets were created in the portal, outside Terraform.
Scope stays on that landing zone. Spend is the cloud invoice. Duration is the pipeline. The figures below are those two instruments, on that estate.
Deterministic structure, then nothing else.
Node pools autoscale, with spot fallback and rightsized pod requests. Azure DevOps YAML is centralized, and dependency caching brought builds to 11 minutes. Portal writes are blocked. Terraform and Azure Policy as Code own the estate.
The same discipline sits under the FinOps stop-loss bundle and under the Assembly Method: attested structure first, unbounded generation later, and only where the remainder is explicit. The 80M token tipping point is the fiscal version of that rule for inference. This paper is the version for the landing zone itself.
Before and after, on one landing zone.
| Plane | Before | After |
|---|---|---|
| AKS node pools | Static 16-node minimum held through off-hours. | Autoscaling, with spot fallback and rightsized pod requests. |
| CI/CD | 40-minute builds. Duplicate SAST on unmerged pull requests. | Centralized Azure DevOps YAML and dependency caching. Builds at 11 minutes. |
| State | 18% of assets created in the portal, outside Terraform. | Portal writes blocked. Terraform and Azure Policy as Code own the estate. |
| Model | The purchase the client asked for first. | Absent from the path. |
32% lower
On that landing zone.
40 min → 11 min
Pipeline time on that estate.
“A model multiplies the architecture underneath it.”
The baseline, then the model
A model multiplies the architecture underneath it. On this landing zone, cloud spend fell 32% before a model entered the path.
Builds went from 40 minutes to 11. Portal writes are blocked. Terraform and Azure Policy as Code own the estate.
Continue with the core engineering offerings or send a contact / B2B inquiry.
Landing zone
Read the invoice before the model
Send the landing-zone invoice and one pipeline's duration. The first question is whether a model belongs in the path yet.
architecture@neplusultra.eu