For data & AI companies
AWS cost optimisation for data and AI companies
Data and AI workloads have legitimately high AWS bills — which is exactly what makes their waste invisible. A five-figure monthly spend hides idle GPUs, running notebooks and untiered datasets that would be screaming anomalies in any other account.
The free scan checks GPU utilisation, notebook uptime and storage tiering across every region — the idle-but-expensive list arrives in 20 minutes.
Run the free scanWhere data & AI AWS bills leak
- →GPUs idle between runs. A p3 or g5 instance left running between training jobs burns more per idle hour than most companies’ entire waste. Training infrastructure should exist while training runs — spot instances, scheduled clusters, or plain stopping when done.
- →Notebooks that never stop. SageMaker notebook instances and Studio apps left running after the experiment ended — the data-science equivalent of leaving the lights on, at ml-instance prices, weeks at a time.
- →Datasets in Standard forever. Training datasets, intermediate artifacts and model checkpoints from every experiment since 2023, all in S3 Standard. Data you touch quarterly belongs in Infrequent Access or Intelligent-Tiering — 40–70% cheaper for the same bytes.
- →Pipelines running for nobody. Scheduled ETL and feature pipelines feeding dashboards nobody opens and models nobody deployed. Data infrastructure accretes consumers slowly and loses them silently — the pipeline runs on.
- →The everything-cluster. One always-on EMR/Spark/Kubernetes cluster sized for the largest job, running 24/7 for jobs that need it four hours a day. Transient, job-scoped clusters exist for exactly this.
Put numbers on it
Frequently asked questions
Our AWS bill is high because we do ML. How do we know what’s legitimate?
By utilisation, not by size. The audit separates "expensive and busy" (legitimate) from "expensive and idle" (waste) using CloudWatch evidence per resource: GPU utilisation on the accelerated instances, job history on the clusters, access patterns on the storage. High spend with high utilisation passes; the report only argues with the idle part.
Should we use spot instances for training?
Usually yes — checkpointed training tolerates interruption, and spot pricing runs 60–90% below on-demand for GPU capacity. It is a bigger engineering change than the audit’s first-line findings (stop the idle, tier the storage), so we sequence it after the free wins.
£499 fixed. Free scan first. 20%+ found or it’s free.
The scan is read-only — a role you create and delete, no keys shared — and shows your estimated monthly saving before anyone talks about money.