Cast AI

Akamai runs infrastructure at a scale where a single percentage point is a budget line. Cast AI cut 40–70% of their cloud spend and returned engineering time alongside it. "An iPhone moment. Two minutes into the integration we saw the cost analytics, and I had an insight I had never had before." - Dekel Shavit, Sr Director of Engineering

𝗔𝗟𝗟𝗘𝗡 𝗗𝗶𝗴𝗶𝘁𝗮𝗹 ran seven models on Amazon SageMaker - three open-source, four custom. • 𝗚𝗣𝗨 𝗶𝗻𝘀𝘁𝗮𝗻𝗰𝗲𝘀 sat underutilized while they paid for full capacity. • 𝗚𝗣𝗨 𝘁𝗶𝗺𝗲-𝘀𝗹𝗶𝗰𝗶𝗻𝗴 put multiple models on one instance. • 𝗟𝗮𝘁𝗲𝗻𝗰𝘆 matched or beat SageMaker. Cost fell more than 70%. • 𝗔𝗻 𝗶𝗱𝗹𝗲 𝗖𝗣𝗨 core costs cents an hour. An idle GPU costs dollars.

Caudalie's team had already done the hard work. ARM on Graviton. 80% of production on Spot. The application tuned on both the code and resource sides. Cast AI still took 40% off their EC2 bill. Hand-tuning gets you to a good configuration once. Automation holds you at the best one as workloads change.

Unstract put Kubernetes on 𝗮𝘂𝘁𝗼𝗽𝗶𝗹𝗼𝘁 and got 𝟴𝟬% 𝗼𝗳 𝗶𝘁𝘀 𝗗𝗲𝘃𝗢𝗽𝘀 𝘁𝗶𝗺𝗲 𝗯𝗮𝗰𝗸. - 𝗥𝗲𝗾𝘂𝗲𝘀𝘁𝘀 are now set from three days of real usage rather than a worst-case guess, with headroom left on the limit. - 𝗠𝗲𝗺𝗼𝗿𝘆 𝗮𝗹𝗲𝗿𝘁𝘀 moved from P0 to P1 because the agent resolves them before they turn critical. - 𝗠𝗼𝗻𝘁𝗵𝗹𝘆 team platform reviews instead of daily.

The usual objection to GPU sharing is isolation. Nobody wants an inference endpoint sharing silicon with a training job that spikes. MIG answers it in hardware. A100, A30 and H100 divide into physically isolated instances, each with dedicated compute cores and its own memory. No noisy neighbours, because there is no shared pool to be noisy in. Cast AI configures the partitioning and places workloads across it automatically, alongside time-slicing and MPS where those fit better.

CPU overprovisioning went from 40% to 69% in a single year. Memory: 79%. Teams are paying for capacity their workloads never even request. We measured 23,000+ production K8s clusters for the 2026 State of Kubernetes Optimization Report - utilization benchmarks by cloud, Spot survival curves by region, and what the top teams do differently. Free download.