Cases from my work: what was built, what was decided, and the figures already published.
Data extracted once to object storage as frozen columnar files. Short single-purpose steps that talk through storage, not to each other. Independent slices run in parallel on short-lived serverless functions on ARM, sized step by step.
An elastic ingestion and processing pipeline on AWS, designed to keep data intact and available at large scale, with the whole infrastructure replicated 1:1 on Azure, where a separate team owned disaster recovery. FinOps and DevOps were part of the design from the start, with infrastructure as code in Terraform, containers on Docker and Kubernetes, and CI/CD. Elasticity and zero interruption were bought with a much larger operational and security surface, which made NIS2 conformance load-bearing.
In the same engagement, a high-performance computing platform on AWS, designed and implemented as code: managed SLURM, FSx for Lustre, MPI and tightly coupled parallel jobs on an auto-scaled EC2 cluster.
Technical director of the engineering entity of a group of about 350 people. Standards, reference integration patterns and a development governance charter with three stated criteria: explicitness, simplicity, maintainability. ADRs made a habit rather than a mandate: teams got the format, the charts and protected time to write them. C4 introduced as the shared vocabulary across domains.
Managed frontier models for reasoning-heavy features, self-hosted lightweight Mistral models where data residency or per-token cost dominates. The trade-off: two runtimes and two security baselines, against lock-in and a token bill nobody could attribute. Cost per feature became a design input, and the reach of each data-processing agreement is decided at the boundary, per capability.
One company per episode, read from the outside: built from public sources, with estimated figures that are not the company's real numbers. I have no relationship with these companies.
From Alan's sandbox problem to a naive agent on AWS at about $206,000 a month at list, cut lever by lever to about $42,000.
From 'pay what you request' to a naive EKS cluster at about $866,000 a month at list, cut lever by lever to about $244,000.
From Doctolib's write wall to a naive Aurora cluster at about $421,000 a month at list, then lever by lever to about $186,000.
From a fleet split three ways to 200 always-on GPU instances on AWS at about $26M a month at list, cut lever by lever to about $12M.
FinOps and cloud architecture, by the day or paid on results.