MinIO AIStor Has Surpassed CoreWeave LOTA Storage Speed

New benchmarks show object storage performance reaching 33.5 GiB/s, potentially removing the need for caching layers.

Updated on Oct. 2, 2026 in Data Centers

Isometric editorial illustration of modular storage drive chassis in a data center, representing high-performance object storage systems. — MinIO's AIStor object storage has recorded performance speeds of 33.5 GiB/s, potentially reducing the need for intermediate caching layers in AI training clusters. AI Illustration. Upload story photo >

Live Poll

Do you believe software efficiency can replace the need for complex hardware caching systems?

MinIO architect Daniel Valdivia published benchmark results demonstrating that MinIO AIStor achieved 33.5 GiB/s of S3 GET performance per node. These tests compare the storage system against CoreWeave LOTA, which recorded 18.4 GiB/s per node in similar testing configurations.

Why it matters

The findings suggest that high-performance object storage architectures may eliminate the need for complex caching systems to keep modern GPU clusters saturated. By moving data efficiently enough to feed compute nodes directly, infrastructure managers could simplify AI training stacks.

MinIO AIStor delivered 33.5 GiB/s per node in throughput using QLC SSDs, significantly exceeding the 18.4 GiB/s per node achieved by the CoreWeave LOTA caching proxy. The CoreWeave tests utilized 20 GPU nodes with 8 GPUs each and 1 TiB of cache per node, supported by dual 100 Gbps network adapters.

The players

MinIO

A company specializing in high-performance object storage software optimized for AI and data lakehouse architectures.

CoreWeave

A cloud provider focused on large-scale GPU compute infrastructure for AI training and rendering workloads.

The details

MinIO AIStor utilizes an erasure-coded object storage architecture running on Solidigm QLC (Quad-Level Cell) flash drives over standard TCP networking. To maintain data consistency, the platform employs MemKV, a shared-nothing key-value store that coordinates storage nodes without a centralized metadata bottleneck. This approach aims to provide sufficient throughput directly from the storage layer, removing the overhead and latency associated with maintaining dedicated caching proxies like LOTA in AI training workflows.

October 2, 2026: MinIO published the performance comparison article.

Data Centers
Sept. 30
CoreWeave Expanded Stack Beyond GPU Compute
Data Centers
Sept. 30
CoreWeave Launched Nvidia Vera Rubin NVL72 System
Data Centers
Sept. 25
NetApp Announced Intent to Acquire PEAK:AIO

The Tech Race

The emergence of high-speed object storage directly challenges the current reliance on the LOTA caching proxy as the industry standard for feeding GPU clusters. By optimizing throughput at the storage layer, this development shifts the competitive focus from caching architecture efficiency back to underlying storage performance.

IT infrastructure teams managing AI workloads may soon reassess the necessity of expensive, complex caching layers in their data pipelines. This approach is most relevant for large-scale GPU deployments currently constrained by storage-to-compute bandwidth bottlenecks.

The takeaway

Direct object storage performance is advancing toward a threshold where traditional caching proxy architectures may become obsolete for AI training. Watch for subsequent industry benchmarks to see if this throughput performance holds under real-world, concurrent GPU-heavy workloads.