ScaleOps

Almost every Kubernetes cluster has at least one pod that can never be evicted, and most teams do not know it is there. New video: the 3 patterns that quietly block Karpenter and Cluster Autoscaler, and how to check for them. Watch below ⬇️

Always-on AI agents have a real cost. Most of it is a GPU sitting idle behind them, 77% of the time on the fleet we measured. Horizontal Pod Autoscaler scales on CPU and memory by design. It has no native GPU signal, so it once collapsed this fleet to one replica while the GPU sat pegged at 100%. Read the blog to see how we killed the idle GPU burn on this fleet, cold starts and all, without adding latency. Link in comments ⬇️

Last month, we wrapped up our strongest quarter to date. Celebrating growth and records is the easy part. The number that tells the real story is ARR per Scaler. We grew the team a lot over the last 12 months, and ARR per Scaler still grew 1.95x. That says something about the market, the product, and mostly - about the people building ScaleOps. AI workloads are everywhere now - inference, LLMs, AI agents, MCP servers, tools. All of it requires heavy compute, and all of it is very bursty. Companies cannot manage their resources manually at that scale. This is creating massive demand for autonomous AI infrastructure management, the problem we have spent four years building for. It drove our growth over the last few quarters, and demand is bigger than ever. We run AI-first as a principle. We research with AI, build with AI, provide value to our customers with AI, and decide with AI. And we have A players who are customer obsessed. AI makes a great team faster. That combination is the whole story. Proud to be building ScaleOps in this era, with this team. Nir Cohen, Taylor Grabus, Zak Blawie ☁☸️, Yarden Weber, Ben Grady, Joey Balázs, Rotem Ben Hamou, Eyal Zilberberg, Adi Steiner, and all the Scalers! This is just the beginning 🚀

Manual GPU tuning doesn't scale with AI workloads that change by the minute. ScaleOps autonomously observes how each AI workload actually uses its GPU, compute and memory, in real time. From there, it automatically assigns the right fractional GPU policy and keeps adjusting allocation as usage shifts. No static slicing. No manual retuning. Just GPUs running at the density your workloads actually need.

Last week we got the whole R&D department together for an evening out, and it reminded us why we love building here: real technical depth, people who genuinely enjoy solving the hard stuff, and a team that knows how to have fun together too. We're growing fast. We currently have 14 open roles across R&D, including: → Backend Engineer → AI Engineer → Infra Engineer If you want to work on complex infrastructure problems (Kubernetes resource management at scale, GPU optimization, real production systems) alongside people like this, take a look. Open roles: https://lnkd.in/dX38DgpW

Island runs one of the most demanding Kubernetes environments on Amazon Web Services (AWS). Thousands of production workloads. Multiple regions. 4-5x growth in a year. Replicas were either too few, causing latency, or too many, wasting resources, with no dynamic way to match demand. With ScaleOps automating workload rightsizing and replica scaling, Island cut Kubernetes vCPU costs by 40%, brought 90% of production under ScaleOps, and runs 1,000+ workloads without manual intervention. Watch the full case study here: https://lnkd.in/d9GvUBBa

Adobe runs 500+ production EKS clusters. Over a million CPUs, thousands of GPUs. All of it now managed autonomously by ScaleOps. In production, with no performance impact, and 38% less compute spend. Case study in the comments.