AIConnect, THE AI DATA PATH

The data path between storage and the GPU.

AIConnect is the AI data path inside Scality ADI. Multiple paths run in parallel so training pulls sustained throughput, inference holds low latency, and retrieval gets the metadata performance it needs.

<div style="display: flex; align-items: center; justify-content: center; gap: 10px;"><img alt="icon-scality" height="30" style="height: auto; max-width: 100%; width: 30px;" src="/hubfs/147481962/Scality%20Web%20Images/icon-scality.webp" width="30" loading="lazy">Explore the agents</div>

WHY ONE PATH THROUGH STORAGE ISN'T ENOUGH

AI workloads don't read storage the same way.

Training pulls sequential bandwidth across petabytes. Inference pulls sub-millisecond reads against KV caches. Retrieval pulls metadata-heavy concurrent reads against billions of objects. Agents pull mixed and bursty. A single path forces every workload to a compromise. The AI data path has to vary with the workload, or the workload pays for it in latency, throughput, or both.

The connector path

Scality ADI delivers multi- TB/sec reads and writes and sub-millisecond latency, including RDMA-enabled S3 where appropriate. Not every AI workload reads the same way, and a single connector path cannot serve every access pattern. AIConnect runs several paths in parallel and routes each workload to the one that fits.

S3 over RDMA

GPU Direct

KV cache connector (sproxyd path)

Workload routing

Ecosystem integrations

Scality ADI integrates seamlessly with the AI infrastructure customers are deploying in production. Certifications, reference architectures, and tested integrations — not vendor compatibility lists.

NVIDIA AI Data Platform & CuObject

WEKA Connector

AI ISV certification

Reference architectures

Workload-aligned access

Each workload type uses the connector path that fits it. Training pulls sequential bandwidth. Inference pulls sub-millisecond KV cache reads. Retrieval pulls metadata-heavy concurrent reads. Agents pull mixed and bursty. The routing is the value, not any single path.

Training: GPU Direct / S3 over RDMA

Retrieval: S3 with object metadata

Inference: KV cache connector

Agents: per-call routing

Scality ADI

Explore Scality ADI

WHERE AIConnect RUNS

Multi-TB/s ingest, sub-millisecond latency, NVIDIA AIDP and CuObject integration, KV cache connector, and workload- aligned routing all ship under ADI as platform capabilities. The AI data path is part of the product, not bolted on.