CoreWeave AI Object Storage

Fully managed S3-compatible object storage from CoreWeave, purpose-built for AI workloads (training datasets, model weights, checkpoints, embedding stores). Architecture separates compute from storage but keeps GPU-local caching tight via **Local Object Transport Accelerator (LOTA)** — a proxy service that runs on the GPU-node hardware, presents an S3 endpoint locally, and uses the node's disks as a tiered cache. Net effect: up to **7 GB/s per GPU** sustained read throughput for AI training pipelines.

Generic S3-compatible object storage was built for general-purpose workloads — high throughput per request but not per-GPU optimized. AI training spends most of its wall-time in dataloader code waiting for objects to arrive; bottlenecking on object-storage egress costs effective utilization (GPU starvation). CoreWeave's bet: co-locate the storage access layer on the GPU node itself via LOTA, exposing S3 API while transparently caching to local NVMe. The 2026 expansion plan extends LOTA acceleration beyond CoreWeave's own cloud into other public clouds and on-prem environments.

Large LLM training pipelines where data movement is the gating factor, checkpoint shuffle stores during long training runs, model-weight distribution across multi-GPU inference fleets, embedding stores for retrieval-augmented training workflows, and any AI workload where per-GPU sustained throughput matters more than raw aggregate capacity.