CoreWeave Extends AI Object Storage Across Regions and Clouds

One dataset, many locations

CoreWeave says the newest features for CoreWeave AI Object Storage let customers use data anywhere in its regions, in other clouds and on premises. According to the company, most organizations have so far had to replicate datasets into every region where workloads run, which raises costs and creates a risk of data divergence.

With the update, a single dataset can be accessed from anywhere in the world with local disk performance, CoreWeave says. The company ties this to its networking backbone, which it describes as combining private interconnects, direct cloud peering and cross-region networking capable of up to 400 gigabits per second.

CoreWeave gives an example: a model can be trained on CoreWeave and then fine-tuned or deployed on another cloud without dataset replication or loss of performance. The company says this portability comes without egress, ingress or request fees.

How the service works

CoreWeave describes AI Object Storage as a fully managed object storage service built to maximize throughput for AI workloads. It uses a distributed architecture that separates compute from storage while keeping latency low, and data is spread across GPU nodes for parallel reads and writes.

Performance relies on the Local Object Transport Accelerator (LOTA), a proxy service that runs on the nodes of GPU clusters. It acts as an S3 endpoint and builds a local cache across the disks on those nodes, staging frequently accessed objects close to the compute. CoreWeave says this removes the need for customers to build separate caching layers. LOTA already runs in every CoreWeave region, and the company says LOTA acceleration for other clouds and on-premises environments is scheduled for early 2026.

The company claims throughput of up to 7 GB/s per GPU, scaling to hundreds of thousands of GPUs, and says performance grows linearly with each added GPU while legacy hyperscalers lose throughput per GPU.

Pricing, reliability and security

Billing is usage-based and automatic. Data is assigned to lower-cost levels according to access frequency:

• Hot: accessed in the last 7 days
• Warm: accessed in the last 7-30 days
• Cold: not accessed in 30+ days

CoreWeave says these levels reduce existing customers' storage costs by more than 75% for typical AI workloads.

On reliability, the company states the service guarantees 99.9% uptime and is designed for eleven nines of durability. Encryption at rest and in transit is standard, and role-based access and SAML/SSO support are included. Observability comes through Grafana dashboards and Prometheus endpoints covering throughput, latency, cache efficiency and error rates. The service is fully S3 compatible.

What it means for AI teams

The update targets a familiar problem: data that is costly to move and tied to where it was first stored. If the claims hold in practice, teams could choose where to run training, fine-tuning or serving based on performance, cost and compliance rather than storage constraints, as CoreWeave puts it.

Teams could also work from a single source of truth instead of several copies, which the company says protects integrity and simplifies workflows. All performance, pricing and cost-saving figures come from CoreWeave itself, and the acceleration for other clouds and on-premises setups was still scheduled for early 2026 in the announcement.