Cloudian has launched its Hyperscale AI Data Platform, an on-premise S3-based storage platform plus artificial intelligence (AI) infrastructure bundle aimed at enterprises that want quick answers from corporate information.
The offer utilises Cloudian object storage plus Nvidia RTX Pro 6000 Blackwell graphics processing units (GPUs) in a retrieval augmented generation (RAG) architecture to power large language model (LLM) functionality that is trained on the mass of corporate data that often goes untapped.
The target use case is to provide natural language querying of corporate data to allow employees to get rapid answers from the data held, which could be about company procedures, data useful for marketing or product development, past code bases, and so on. Cloudian emphasises that the product works fully on-premise and is “air-gapped” to ensure the security of the organisation’s data.
It comprises three nodes of S3 object storage, in this case on-premise, and connected using S3 over remote direct memory access (RDMA), developed with Nvidia. This allows for rapid connectivity between storage nodes, using RDMA, which was originally developed to allow data to move from the memory of one server to another for high-throughput, low-latency operations while not hitting central processing unit (CPU) resources.
S3 over RDMA leverages this approach to cut latency by bypassing the TCP/IP stack. This aims to address the bottlenecks that can occur between storage nodes during AI processing, which are a key constraint on AI performance.
Sitting above this, but at the heart of the platform, is a so-called billion-scale vector database. Vector databases have emerged as core to AI as it has taken the fore. As data is ingested into an AI system, its characteristics are given multiple numeric values. These values can then be computed upon to calculate similarity, context, and to give some semblance of meaning.
In Cloudian’s Hyperscale AI Data Platform, any new information can be ingested without the need to retrain the entire corpus, while the architecture also supports images and structured data as well as text in unstructured data at which the product is chiefly targeted.
Cloudian is one of a number of suppliers that offer enterprise object store products. There are, in fact, 22 object storage players in analyst house GigaOm’s 2025 Object Storage Radar. A few of these have some kind of play into object storage platforms aimed at AI use cases, with Cloudian ranked among the most innovative among them.
Also present in that space with object storage, RAG and vector database capability are specialists such as Scality and Minio, and general storage players like Pure Storage and NetApp.
Cloudian’s object storage comes under its HyperStore family. It is native S3, but also allows for SMB and NFS file access. Hyperstore nodes come in a range of spinning disk HDD models plus an all-flash option with TLC NVMe drives.
Cloudian’s Hyperscale AI Data Platform uses the Llama 3.2-3B-Instruct LLM. Meanwhile, its four Nvidia GPUs are dedicated to different phases in the workload, namely LLM inferencing, vector database operations, re-ranking and relevance, and one shared for vector embedding and other functions.
Users get an easy-to-use graphical user interface that allows them to ask questions in natural language and then refine them, just as they would using any popular LLM.
Target use cases include enterprise knowledge mining, secure document intelligence, video content analysis, and building data lineage and audit trails for compliance and governance purposes.
