Global advertising holding (Top 5 worldwide)
Multi-agency platform on Amazon EKS with Karpenter just-in-time GPU scaling. 76% cost reduction per generation versus a fixed fleet.
The customer's portfolio of agencies was managing AI creative infrastructure independently — duplicating GPU spend and fragmenting tooling. Fixed EC2 GPU fleets projected to exceed $100K/month at target adoption, well before reaching meaningful utilisation. Third-party creative tools needed programmatic API access that didn't exist. They needed one centralised, multi-tenant platform with isolated environments per agency and intelligent scaling that eliminated idle GPU spend.
A multi-tenant containerised platform on Amazon EKS with Karpenter providing autonomous just-in-time GPU provisioning across three node pools — g5 for standard workloads, g6e.12xlarge L40S for high-quality image generation, and p4de.24xlarge A100 with NVIDIA MIG partitioning for video. SQS distributes work; WebSocket delivers real-time progress to the UI. Scale-to-zero eliminates GPU costs during off-hours. API Gateway plus Lambda exposes workflow execution to third-party tools. Namespace-level isolation with Kubernetes NetworkPolicies and per-agency RBAC. CI/CD via GitHub Actions and Helm.
Per-generation cost dropped from $2.30 (fixed fleet) to $0.55 (Karpenter dynamic) — a 76% reduction. GPU cold-start under 50 seconds from submission to inference readiness. Scale-to-zero gives a true zero GPU spend during non-working hours. NVIDIA MIG partitioning lets multiple users share a single A100 concurrently. Platform spend projected at $45K/month at current adoption, against $100K+ for the fixed-fleet alternative.
Under the hood
Customer name redacted at the customer’s request. Numbers, services, and architecture are unchanged.
Next case study
Global advertising network (Top 5 worldwide)