Local Object Transport Accelerator (LOTA)
An AI-native caching/transport proxy that runs on GPU/CPU nodes and presents a local S3 endpoint — serving hot data from node-local NVMe while pushing cold data to object storage, giving compute parallel S3 reads without cross-region egress penalties.
Summary
An AI-native caching/transport proxy that runs on GPU/CPU nodes and presents a local S3 endpoint — serving hot data from node-local NVMe while pushing cold data to object storage, giving compute parallel S3 reads without cross-region egress penalties.
The economics-meets-locality layer. At tens of thousands of GPUs, traditional object storage flattens under parallel load; LOTA colocates cache + transport on the node so a unified global dataset is reachable everywhere without replication drift.
- It is a proxy/cache, not a new storage system — the durable copy still lives in object storage.
- The cost win (up to ~75%) comes from automated tiering + egress avoidance, not cheaper bytes.
- LOTA
depends_onCoreWeave AI Object Storage — ships as part of that platform - LOTA
solvesCloud AI Storage Price Inversion — cuts egress and storage cost - LOTA
acts_asCache-Fronted Object Storage;optimizes_forInference Locality
Definition
An AI-native caching and object-transport proxy that runs directly on GPU/CPU cluster nodes, presenting a local S3 endpoint. It serves frequently accessed (hot/warm) data from local NVMe while transparently pushing cold data back to persistent object storage, giving compute highly parallel S3-compatible reads without cross-region egress penalties.
At tens of thousands of GPUs, traditional object storage flattens under parallel load and cross-region synchronization incurs heavy egress cost and latency. LOTA collapses that by colocating a transport+cache proxy on the node, so a unified global dataset is reachable everywhere without replication drift or duplication.
High-throughput training-data and checkpoint access on large GPU clusters, multi-region/multi-cloud dataset access without sync, egress-cost reduction, hot/warm tiering at the node edge.
Recent developments
- Reached GA in 1H 2026 on CoreWeave AI Object Storage. Runs on GPU nodes as a distributed proxy, delivering up to ~7 GB/s per GPU with throughput scaling linearly as the cluster grows. Per CoreWeave — AI Storage Without Limits and CoreWeave Docs — About LOTA.
- Usage-based tiering cuts storage cost up to ~75%. Automated hot/cold movement removes manual tiering overhead and unifies datasets across regions/clouds without duplication. Per CoreWeave Docs — About LOTA.
Connections 5
Outbound 5
scoped_to1depends_on1optimizes_for1acts_as1Resources 2
CoreWeave's reference for LOTA — the node-local S3 proxy, per-GPU throughput numbers, automated tiering, and the egress/cost-reduction mechanics.
Context on CoreWeave AI Object Storage and how LOTA scales linearly with cluster size for multi-region datasets.