# Data that keeps up with reasoning

## Executive summary

CoreWeave addresses data bottlenecks in large-scale AI training and inference by offering a specialized storage portfolio. The core solution, CoreWeave AI Object Storage, is designed for massive parallel throughput, achieving up to 7 GB/s per GPU. Key differentiators include S3 compatibility, no egress/ingress/request fees, and the patented Local Object Transport Accelerator (LOTA) which provides global, high-speed caching directly to GPU nodes, ensuring maximum GPU utilization and lowering total cost of ownership (TCO).

## Key takeaways

- High-Performance, Scalable Object Storage: CoreWeave AI Object Storage provides massive parallel throughput, scaling with the addition of GPU nodes, and supports up to 7 GB/s per GPU of throughput. It is S3-compatible and designed specifically for AI workloads.
- Global Data Accessibility and Cost Control: The service has no egress, ingress, or request fees, making it ideal for multicloud environments. Furthermore, it uses automated, usage-based billing (Hot, Warm, Cold tiers) to reward users for inactive data without requiring manual tiering or data movement.
- Local Object Transport Accelerator (LOTA): LOTA is a proxy running on GPU nodes that creates a global cache across all nodes' NVMe drives. This accelerates data access by serving reads and writes from the local cache, significantly boosting performance and enabling fast cross-region data access.
- Optimized GPU Utilization: By ensuring data is fed to the GPUs as quickly as possible, CoreWeave AI Object Storage maximizes GPU utilization, which is critical for lowering the overall total cost of ownership for large compute clusters.

## Technical details

- CoreWeave AI Object Storage: Primary solution for large-scale training and inference. It is S3-compatible, offers high throughput, and is purpose-built to keep storage pace with compute.
- LOTA (Local Object Transport Accelerator): A patented proxy running on GPU nodes (Kubernetes pod) that creates a global cache across all nodes' NVMe drives. It allows S3 applications to talk directly to the proxy, bypassing traditional gateway layers.
- Billing and Data Tiers: Automated usage-based billing with three tiers (Hot: last 7 days; Warm: 7-30 days; Cold: 30+ days). Billing is automatic, requiring no programmatic changes or data movement.
- Cross-Region/Cross-Cloud Data Management: LOTA extends performance across CoreWeave regions and even to other clouds (e.g., GCP, OCI) via Helmcharts, allowing data to be accessed and written across locations without replication or incurring fees.

## Practical implications

- Build engineers can design ML pipelines knowing that data bottlenecks are mitigated by specialized, high-throughput storage.
- The ability to access data across multiple cloud regions/providers (multicloud) simplifies deployment and data residency concerns.
- Automated billing and high efficiency directly contribute to lower Total Cost of Ownership (TCO) for compute-intensive AI workloads.
- The S3 compatibility and deep observability (via Grafana) allow for easy integration into existing MLOps and CI/CD workflows.

## Topics

AI/ML Workloads, Object Storage, High-Performance Computing (HPC), Cloud Infrastructure, Data Streaming, CoreWeave AI Object Storage, LOTA (Local Object Transport Accelerator), CoreWeave Cloud

Source: https://www.youtube.com/watch?v=fLcBZPiSFcY