CoreWeave makes AI Object Storage generally available at NVIDIA GTC
CoreWeave has made CoreWeave AI Object Storage generally available, a managed, S3-compatible object storage service built for AI training and inference. The company announced the launch at NVIDIA GTC.
- CoreWeave AI Object Storage is generally available, announced at NVIDIA GTC.
- The S3-compatible service targets AI training and inference, with up to 2 GB/s per GPU, according to CoreWeave.
- LOTA caches or prestages data on local NVMe disks of GPU nodes to cut network latency.
- CoreWeave says there are no egress, per-request, or hidden fees.
- It integrates with CKS, SAML/SSO, Terraform and managed Grafana dashboards.
What CoreWeave is launching
According to CoreWeave, AI Object Storage is designed for GPU-intensive model training and offers exabyte-scale capacity. It joins the company's existing storage lineup, which includes Distributed File Storage and Dedicated Storage.
CoreWeave argues that traditional cloud storage falls short as models grow in size and complexity. It says the new service was engineered from the ground up to meet those performance demands, and that traditional object storage systems often create bottlenecks in data-intensive AI workflows.
The service is built to work with CoreWeave's NVIDIA GPU compute clusters. The company states it supports performance levels of up to 2 GB/s per GPU and scales to hundreds of thousands of GPUs. It also says the service fully supports the latest NVIDIA GPU architectures, including Blackwell and Hopper.
LOTA caching and pricing
A central part of the product is the Local Object Transport Accelerator, or LOTA. According to CoreWeave, it caches frequently used datasets and/or prestages data directly on the local NVMe disks of GPU nodes. The company says this reduces network latency and improves training speeds.
CoreWeave also says the caching technology can operate without additional tools or complex configurations, so AI teams can concentrate on building and refining their models.
On pricing, the company lists the following:
- No egress, per-request, or hidden fees, with customers paying only for the storage they use.
- Capacity for trillions of objects and exabytes of data, with performance that grows with storage needs.
- Encryption at rest and in transit, role-based access control, audit logging, and lifecycle policies.
Integration with the CoreWeave platform
CoreWeave says AI Object Storage is built to fit into its cloud ecosystem. It supports the platform's SAML/SSO authentication for user management and identity federation. LOTA is integrated with CoreWeave Kubernetes Service (CKS), which the company says allows tight integration between CKS applications and the caching layer.
Developers can manage storage with the CoreWeave Terraform provider, which enables automation and infrastructure-as-code workflows, and deploy storage resources alongside GPU compute clusters.
A managed Grafana dashboard is available from the CoreWeave Console. It shows requests per second, throughput or bandwidth usage, and request duration. Data can be grouped or filtered by bucket, region, operation type, or response code, which CoreWeave says helps teams understand usage patterns, spot issues and troubleshoot.
What it means for AI teams
For teams already running training or inference on CoreWeave GPUs, the launch adds an object storage option that is S3-compatible and tied to the same infrastructure. The absence of egress and per-request fees, as described by the company, makes the pricing model a storage-only charge.
The performance figures, including up to 2 GB/s per GPU, are CoreWeave's own claims, and the source does not include independent benchmarks. The company also points to its wider infrastructure design, which it says unlocks up to 20% higher GPU performance.
