TechPrufer
Get started
Coreweave
Coreweave logoCoreweave · Feature trail

AI Object Storage Cost Optimization

15 posts · May 2026 – Sep 2026

What this trail covers

CoreWeave AI Object Storage (CAIOS) is generally available (GA) with enhanced performance and cost-efficiency for AI workloads. It integrates seamlessly with CoreWeave's NVIDIA GPU compute clusters, supporting performance up to 7 GB/s per GPU and scaling to hundreds of thousands of GPUs. The Local Object Transport Accelerator (LOTA™) caches frequently used data, achieving 7 GB/s per GPU throughput. New automated usage-based pricing levels (Hot, Warm, Cold) reduce storage costs by over 75% for ty. This post details configuration for LOTA caching, pre-staging, cross-region writes, and Archive tiering, providing Python examples for faster reads and writes at lower cost.

All Coreweave trails
  1. CoreWeave AI Object Storage Setup: LOTA, Writes, ArchiveThis post details the configuration of CoreWeave AI Object Storage (CAIOS) features: LOTA caching for accelerated reads, cross-region write acceleration, and the Archive storage tier for cost optimization. It provides Python (boto3) and CLI examples for pre-staging LOTA cache with `head_object` calls, enabling cross-region writes via `upload_file` to a bucket, and writing directly to Archive using `ExtraArgs={'StorageClass': 'STANDARD_IA'}` or enabling automated archiving with `cwic cwobject bucket update`. It also explains how to read from the Archive tier and copy objects to the STANDARD class.
  2. CoreWeave AI Object Storage: Cross-region writes, ArchiveIntroduces cross-region write acceleration for CoreWeave AI Object Storage, enabling writes to remote buckets at local latency by writing durably to the local region first and then migrating data to the remote region in the background. Also introduces a new Archive storage class for infrequently accessed data, with no retrieval, cache, request, early-deletion, tiering, or egress fees, and a longer time to first byte compared to other tiers. These features aim to reduce operational overhead and cost by allowing compute to be placed where available without data relocation complexities.
  3. Video Captioning at Scale: 600 TB in 95 Minutes | CoreWeave BlogThis post details the successful execution of a large-scale video captioning pipeline processing 600 TB of data in 95 minutes on CoreWeave's infrastructure, leveraging Anyscale's Ray Data. It highlights the performance of CoreWeave AI Object Storage (CAIOS) in handling massive data throughput (40 GB/s writes, 110-120 GB/s reads) and low latency during this operation. The post contrasts the efficiency of Ray Data over Ray Core for GPU utilization and throughput, showcasing how the combined infrastructure enabled rapid deployment from account signup to production job in under 24 hours. It also emphasizes the role of CoreWeave's managed Kubernetes service (CKS) and Mission Control in providing a ready-to-use, healthy AI infrastructure.
  4. AI storage and LLMs: 4 critical needs to look out forThis post details the critical needs for AI storage systems supporting LLMs: fast data access, quick recovery and resiliency, scalability, and airtight security. It highlights CoreWeave AI Object Storage's features, including the Local Object Transport Accelerator (LOTA) for direct GPU-to-data paths, performance metrics (2 GB/s per GPU, 25 GB/s per 1 PB reserved storage, 5000 RPS per account), 99.9% uptime, and eleven nines of durability. It also emphasizes security measures like encryption at rest and in transit, and identity access management with MFA and role-based access controls.
  5. AI Storage Without Limits: Exploring the Latest CoreWeave AI Object Storage Expansions | BlogThis post details the expansion of CoreWeave AI Object Storage (CAIOS) with cross-region and multi-cloud flexibility. It highlights that a single dataset can now be accessed seamlessly from anywhere in the world, including other clouds and on-premises environments, without data replication or egress fees. This is enabled by CoreWeave's networking backbone and LOTA technology, which will expand to other clouds and on-premises environments in early 2026. The post reiterates the performance benefits of LOTA, achieving up to 7 GB/s per GPU, and the cost savings through automated usage-based billing (Hot, Warm, Cold tiers).
  7. Break Free From Hyperscaler Storage Constraints | CoreWeaveThis post details how CoreWeave's AI Object Storage addresses hyperscaler storage constraints for AI workloads by offering a high-throughput, low-latency, flat, unified design with zero egress fees and adaptive pricing. It highlights the architectural differences from conventional storage, emphasizing the purpose-built nature for GPU performance with an InfiniBand backbone and LOTA caching, a unified data plane across regions, and predictable economics. The post also introduces the Zero Egress Migration (0EM) program to facilitate data migration.
  8. CAIOS Achieves 7+ GB/s per GPU on NVIDIA Blackwell Ultra | CoreWeaveThis post details new benchmark results for CoreWeave AI Object Storage (CAIOS) on NVIDIA Blackwell Ultra GPU nodes, achieving an average throughput of 7+ GB/s per GPU. This represents a >3x improvement over previous H100 benchmarks. The gains are attributed to: 1) fewer GPUs per node (4 on Blackwell Ultra vs. 8 on H100), doubling per-GPU throughput; 2) the use of NVIDIA Quantum InfiniBand transport, which improved throughput by approximately 50% compared to Ethernet; and 3) LOTA pipeline optimizations, resulting in an additional ~17% performance improvement. The benchmark setup used the Warp S3 benchmarking tool on CoreWeave Kubernetes Service (CKS) with 16 Blackwell Ultra nodes, testing both Ethernet (TCP) and RDMA (InfiniBand) transports.
  9. CoreWeave AI Object Storage Delivers 2+ GB/s per GPUThis post details benchmark results for CoreWeave AI Object Storage (CAIOS) demonstrating read throughput of over 2 GB/s per GPU. It highlights the Local Object Transport Accelerator (LOTA™) feature, which caches data on local NVMe disks within GPU nodes. The benchmark methodology involved deploying a test application on 20 NVIDIA H200 GPU nodes using the Warp S3 benchmarking tool, simulating read and write operations to the cwlota.com endpoint. The results show that after initial cache population, the system achieved an aggregate read throughput of 368 GiB/s, averaging 18.4 GiB/s per node or 2.3 GiB/s per GPU. The post emphasizes that this performance is expected to scale to any number of GPU nodes and GPUs due to its reliance on local storage performance.
  10. CoreWeave Launches AI Object Storage: Optimized for AI Workloads | CoreWeaveThis post announces the General Availability (GA) of CoreWeave AI Object Storage, a managed object storage service optimized for AI training and inference. It highlights performance up to 2 GB/s per GPU, scalability to exabytes, and integration with NVIDIA GPUs. Key features include the Local Object Transport Accelerator (LOTA™) for caching and reduced latency, S3 compatibility, and enterprise-grade security. The post also details seamless integration with CoreWeave's ecosystem, including SAML/SSO, managed Grafana dashboards for observability, Kubernetes integration via CKS, and management via the CoreWeave Terraform provider.
  11. Expanding CoreWeave AI Object Storage: Unified Dataset, Lower Costs, and Unmatched ScaleThis post details the expansion of CoreWeave AI Object Storage to include unified global datasets accessible across CoreWeave regions, other public clouds, and on-premises environments. It highlights the LOTA technology for accelerating performance, delivering 7 GB/s per GPU throughput. Additionally, it introduces new automated usage-based billing tiers (Hot, Warm, Cold) that reduce storage costs by over 75% for typical AI workloads, eliminating manual data movement and punitive access fees. The post also announces the general availability of LOTA acceleration cross-regionally and expansion to multi-cloud/hybrid environments in early 2026.
  12. Introducing our Zero Egress Migration program | CoreWeave BlogIntroduces the Zero Egress Migration ([0]EM) program, a service to facilitate large-scale data transfers from other cloud providers to CoreWeave, covering egress fees from the source cloud. The program handles orchestration, end-to-end checksum validation, and provides a real-time migration dashboard. It emphasizes that once data is in CoreWeave AI Object Storage, there are no egress fees charged by CoreWeave. This program aims to remove data mobility as a constraint for AI teams by eliminating egress costs and simplifying the migration process for petabyte-scale datasets.
  13. Lower AI Storage Costs by 75% | CoreWeaveThis post introduces automated usage-based billing levels (Hot, Warm, Cold) for CoreWeave AI Object Storage, which automatically distinguish and lower the price for inactive data without slowing reads or changing pipelines. It explains that data moves from Hot to Warm after seven days of inactivity and from Warm to Cold after more than 30 days, all without manual management or rehydration delay. The post also includes a competitive cost analysis comparing CoreWeave AI Object Storage to AWS S3 Standard, AWS S3 Express, GCP GCS Standard, and Azure Blob, Hot LRS, highlighting CoreWeave's lack of egress, per-request, retrieval, or tiering fees and consistent performance.
  14. Optimize High-Performance Computing Storage | CoreWeaveThis post provides tactical strategies for optimizing storage performance for ML workloads on CoreWeave Cloud, building upon previous benchmarking. It details practical techniques including: implementing multi-level buffering and caching with considerations for system resources; effective prefetching strategies to keep GPUs fed without memory contention; optimizing data format (TFRecord, WebDataset) and organization (sharding) for efficient loading; pipeline optimization to overlap I/O, decompression, preprocessing, and GPU transfer; memory management and distributed coordination for cache contents and prefetch operations across clusters; and monitoring key metrics (I/O wait, throughput, latency, cache hit rates) and implementing error handling for reliable, consistent performance.
  15. Performant, Flexible Object Storage Solutions | CoreWeaveThis post announces the general availability of CoreWeave Object Storage, detailing its features, pricing, and benefits. Key technical contributions include the introduction of S3 compatible API with SSE, an NVMe accelerated option for high throughput, independent regional clusters for bucket names, and the Accelerated Cache for Object Storage utilizing NVMe-backed regional caches for low-latency data access. It also highlights cost-effectiveness with simple pricing and no hidden fees.
  16. Storage Benchmarking: Distributed File Storage | CoreWeaveThis post details CoreWeave's distributed file storage system for model training, highlighting its architecture (VAST Data's disaggregated two-tier system with CNodes and DNodes) and performance characteristics. It presents benchmark results using fio, demonstrating sustained 1 GiB/s per GPU throughput and peak read/write rates of 70 GiB/s and 50 GiB/s respectively for simulated AI training workloads. The post emphasizes how this storage solution prevents GPU underutilization by ensuring storage does not create bottlenecks, evidenced by sustained 100% GPU core utilization.