Aranya Inc., a startup that says it can transform raw bare-metal servers into custom, production-ready graphics processing unit clusters for artificial intelligence inference in less than two days, launched today with $11 million in funding.

The company says it’s addressing a critical bottleneck in the AI infrastructure market. Founded last year, Aranya says it has already been tasked with managing more than $500 million worth of GPU hardware for leading AI inference providers and data centers.

Its flagship technology is an open-source engine called clusterdOS, built on top of Kubernetes, that provides a framework for deploying and maintaining infrastructure through declarative configuration files. It also automates the orchestration of containerized workloads, networking and storage resources across various hardware environments.

“ClusterdOS is built for organization-scale computing, not for managing individual applications or isolated workloads,” Christian Bhatia Ondaatje, co-founder and CEO of Aranya, said in a written response to questions about the platform’s capabilities.

Aranya said it addresses Kubernetes’ shortcomings in two critical areas. While Kubernetes can reschedule a pod off a failing node, it does not diagnose the root cause. “ClusterdOS discovers, diagnoses and resolves the hardware itself,” Ondaatje said.

That includes GPU thermal events, error-correcting code errors and networking faults. The engine is also not Kubernetes-exclusive. It also encompasses virtual machine-based and the Simple Linux Utility for Resource Management job schedules and workload manager. “Real GPU infrastructure is heterogeneous and the OS layer has to absorb that,” Ondaatje said.

The company promises it can transform any quantity of racked hardware into a cohesive, self-healing cluster within 48 hours, including custom storage and networking. “We’ve matched that timeline repeatedly, and no other operator does it consistently for custom architecture,” Ondaatje said.

ClusterdOS provides federated control across multiple clusters. The system is designed to autonomously detect and remediate issues before they escalate, using agents that “live inside the cluster continuously and adapt to what that cluster needs,” rather than executing pre-scripted jobs, Ondaatje said.

Strict controls

To prevent accidental disruption, the OS operates under strict permission controls. “The multicluster OS cannot see or access anything the organization hasn’t explicitly granted it,” he said. “It’s governed by strict multitenancy and role-based permissions at the OS level. Access is granted, not assumed.”

Aranya is also introducing a natural-language interface that allows engineers to manage infrastructure using plain commands, such as spinning up inference endpoints or adding nodes, without touching configuration files. The system respects existing user permissions to ensure that actions are both secure and auditable. “A request in Slack resolves to the same permitted action set the engineer already has,” said Ondaatje.

The company is also claiming significant operational improvements from the use of its software. The most notable are what Ondaatje called “dramatic reductions in downtime at both the cluster and datacenter scale, achieved through proactive symptom detection.

The federated architecture also allows teams to manage all their clusters as a unified fleet, while built-in utilization monitoring helps rightsize workloads and optimize costs. Aranya’s system also enables customers to recycle idle compute, further reducing operating expenses.

With the fresh capital, Aranya plans to expand its engineering and sales/marketing teams and launch a full multicluster interface. The round includes a $9 million seed led by First Round Capital Holdings L.P. with participation from BoxGroup Ventures LLC, Vermilion Cliffs Ventures LLC and Asylum Ventures and a $2 million pre-seed round led by Asylum Ventures.

Image: SiliconANGLE/Microsoft Designer