Aranya raises $11M to build an operating layer for bare-metal GPU clusters

Why it matters

GPU availability solves only the procurement problem. Aranya is betting that the next valuable infrastructure layer will turn heterogeneous hardware into reliable capacity without every AI operator building its own platform team.

Christian Bhatia Ondaatje, Aryamika Bhatia Ondaatje, and Sasivarnan Kanaghasalam Sathyapriya announced $11M in funding for Aranya on September 1, giving their year-old San Francisco operation fresh capital to turn racks of GPU servers into production clusters in under 48 hours.

The financing combines a $9M seed led by First Round Capital with a previously raised $2M pre-seed led by Asylum Ventures, according to Aranya's announcement. BoxGroup, Vermilion Cliffs Ventures and Asylum joined the seed. Founder Collective, Parable VC and Uncommon Ventures participated in the pre-seed.

A securities filing reports $11.3M sold to 11 investors, compared with the company's rounded $11M announcement. Aranya has not disclosed a valuation.

SiliconANGLE reported that the funding will support hiring and the rollout of an AI interface for operating infrastructure across multiple clusters. Engineers would be able to provision inference endpoints, add nodes and diagnose failures through plain-language instructions, according to Aranya.

A fourth compute startup

Aranya is the latest expression of Christian Bhatia Ondaatje's long-running fixation on GPU infrastructure. The Harvard graduate has described Aranya as his fourth compute startup and sixth compute venture overall. His earlier work included the Wolfepack external-GPU project, peer-to-peer GPU cloud Squire, an early engineering role at Crusoe, personal data center project Atoka and a founding platform engineering role at Hyperbolic.

That history gives Aranya a credible founder-market fit for an unusually unforgiving job. GPU clusters fail across hardware, drivers, networking, storage and orchestration, and the expensive part is often the engineering labor required to make those layers behave as one production system.

Aryamika Bhatia Ondaatje came to Aranya from operating and business roles. Her public profile lists MIT Sloan and earlier venture capital experience, and she joined robotics developer Cobot as chief of staff before founding Aranya. Sasivarnan Kanaghasalam Sathyapriya earned an MBA from MIT Sloan in 2025. Before Aranya, he worked on distributing inference jobs across otherwise idle GPUs in electric-vehicle fleets.

Sathyapriya has said Ondaatje's proposal for a multicluster operating-system layer matched the underused-compute problem he had encountered in automotive hardware. The founders are applying that thesis to data centers and AI operators with far more capital tied up in accelerators.

The gap between a rack and a working cluster

Buying or renting bare-metal GPU servers leaves an AI operator with substantial work before a model can serve production traffic. Drivers must match the hardware. Networking and storage have to sustain the workload. Kubernetes, virtual machines or Slurm need to be configured, monitored and upgraded. Hardware faults still arrive after deployment, usually at inconvenient hours.

Aranya's clusterdOS packages that work into an open-source, GitOps-native engine built on Kubernetes. Its published stack includes ArgoCD, Ceph, Cilium, vLLM, Slurm and operators for Nvidia and AMD GPUs. Customers can run the GPLv3-or-later software themselves or pay Aranya to design, deploy and operate the infrastructure.

That distinction matters. Aranya is selling an operating model alongside software. The service covers architecture, deployment and 24-hour operations, while the multicluster layer coordinates infrastructure across locations. Aranya positions that approach against generic managed Kubernetes products, which supply a cluster while leaving workload-specific GPU scheduling, networking and storage to the customer.

Baseten and Hydra Host are named on Aranya's website as customers or partners. Their testimonials are hosted by Aranya and should be read as customer endorsements rather than independent performance tests.

The numbers come from Aranya and its customers

Aranya says it already manages more than $500M worth of GPU hardware, an unusually large figure for a business founded in 2025. Aranya has not supplied the hardware valuation method, customer count or revenue attached to that footprint, and the figure has not been independently audited.

A July 14 case study offers a narrower view of the operating claims. Aranya and Hydra Host say their joint deployment spans more than 1,700 GPUs for an AI inference customer, reaches production in 24 to 48 hours and maintains 99.99% Kubernetes control-plane uptime. The case study also claims a 90% reduction in issue identification and resolution time against the customer's prior baseline.

Those results come from Aranya and Hydra Host, with no public third-party benchmark. They still show where the founders are concentrating: customers that already have access to GPUs and need a small infrastructure team to keep those machines productive.

Aranya sells software and takes the pager

The $11M round gives the founders room to expand both sides of a demanding model. Aranya plans to expand its engineering and sales and marketing teams. The natural-language management interface adds another product surface above clusterdOS.

The hybrid approach can accelerate adoption because customers receive a working cluster rather than another tool to integrate. It also creates a scaling test. Managed infrastructure requires people, customer-specific architecture and on-call coverage. Aranya's software will need to absorb more of that labor as deployments grow if the economics are to resemble a software platform rather than an infrastructure consultancy.

The immediate job is practical: take expensive hardware that has already been delivered, get it into production quickly, and keep its owners from assembling another platform team just to make the rack work.