Aranya

A lot of this week's datacenter news comes back to one date: when the power turns on. But for operators who already have power, there’s a second date that matters just as much: the day the GPUs are online and actually usable. The industry norm from racked servers to a production-ready cluster is weeks, and until then, those GPUs aren't earning. We do it within 48 hours. The only requirement? Hardware is racked. We handle the rest: Bare metal provisioning, GPU orchestration, CPUs, networking, and storage. From there we build the architecture the tenant needs, whether that's Kubernetes, Slurm, VMs, or storage, and hand it over production-ready. Power decides when you can build. The cluster decides when you can earn.

There’s a huge gap between what data centers offer and what AI workloads really need.

Many neoclouds run KubeVirt under the hood. What separates fast from functional isn't the VM layer, but whether you can get GPUs through it cleanly. VFIO passthrough: direct PCIe access from the VM to the GPU, hypervisor completely off the data path. That's what we shipped in clusterdOS v0.3.17. clusterdos-kubevirt deploys KubeVirt v1.8.2 + CDI v1.65.0 as an ArgoCD gitapp. Enable it in your metarepo values and your cluster gets GPU passthrough to H100s, B200s, and B300s, plus live VM migration between nodes and declarative disk lifecycle via CDI. cert-manager is already there. Operators have memory limits set. CDI clone deadlock is patched. Upgrade tooling is automated. Phase 1 of the HyperConverged stack. That used to be weeks of custom configuration per cluster. Now it's a values flag in your metarepo. This is infrastructure complexity moving into the platform.

GPU costs are hot. But what about utilization? Windsurf cut GPU costs 50%. Cartesia tripled their cluster size. Odyssey ran infrastructure at up to 4.4x lower cost than hyperscalers. None of that came from finding cheaper GPUs. It came from infrastructure designed around the actual workload, not general purpose compute retrofitted for AI. The dirty secret of AI infra: most teams are paying for clusters they're not fully using. The waste is at the architecture level, not the billing level. The teams winning on cost aren't squeezing margins. They're building differently. What are you actually optimizing: the billing line or the architecture?

Fast Company just named Aranya as one of The Next Big Things in Tech! Big thanks to the team and everyone who's helped us get here. Link to the full article in the comments ↓

Aranya is a finalist for Best AI Infrastructure Solution at the 2026 A.I. Awards! We're in the running with WEKA, Cloudera, and Experian. So give us a call if you need custom clusters within 48 hours. Winners announced October 13.