G2

Mukul K.

AI Engineer

Mid-Market (51-1000 emp.)

10/4/2026

Business partner of the seller or seller's competitor, not included in G2 scores.

"Automatically Cut Kubernetes Costs Without Sacrificing Speed"

4.5/5

What do you like best about Cast AI?

Automatically cuts our Kubernetes costs without hurting speed. Review collected by and hosted on G2.com.

What do you dislike about Cast AI?

The UI is getting crowded. With security scanning, rightsizing, and spend forecasting added, finding a simple autoscaler setting can take three menus, and the docs seem to lag behind new features. I've had to contact support a few times, though they do respond. Node Templates also have a steep learning curve. You need to understand tolerations and node affinities to keep sensitive workloads like Redis and API gateways off spot nodes. Early on, a rebalance during a traffic spike briefly slowed our Retool dashboard. The savings-based pricing is fair to me as an engineer, but it's harder to explain to finance than a fixed license. Review collected by and hosted on G2.com.

What problems is Cast AI solving and how is that benefiting you?

Our biggest problem was a Kubernetes bill we couldn't explain. We run self-hosted inference, vector databases, and background jobs behind n8n and Retool, and I over-provisioned out of fear of latency spikes and cold starts. High-memory nodes and GPUs ran around the clock while sitting idle most of the time. Tuning requests and limits by hand was painful, and the native cluster autoscaler was too slow.

CAST AI fixed this in three ways. First, it showed us the waste. Our GPU utilization was in the single digits, and the projected savings were 40-50%. Second, it made spot instances usable. Batch document processing, transcription, and embedding jobs now run on cheap spot nodes, and it drains nodes gracefully before interruptions, so we don't get failed jobs or lost requests. Third, its bin packing keeps consolidating workloads onto the cheapest nodes, so we're not paying for idle capacity.

The result is lower compute costs, and I get my time back. I no longer babysit Grafana dashboards or explain why our GCP bill doubled after deploying a new model. Our real-time inference endpoints stay responsive too, as long as pod disruption budgets are set up properly. Review collected by and hosted on G2.com.