MinIO: you don’t need an object cache. You need a faster object store.

MinIO architect Daniel Valdivia published a blog post arguing that AI operators do not need object caches if their object store is fast enough, citing a CoreWeave Warp benchmark in which MinIO AIStor on Solidigm QLC SSDs reached 33.5 GiB/s per node (36 GB/s) with no caching versus 18.4 GiB/s per node (19.8 GB/s) for CoreWeave's LOTA caching proxy across 20 GPU nodes with 8 GPUs each and 1 TiB of LOTA cache per node. Valdivia said "An erasure-coded object store on capacity-optimized QLC drives, running over plain TCP, delivers more throughput per node than a warm cache on local NVMe," and MinIO noted object caches and inference KV caches are distinct, with MinIO supporting KV caching through its MemKV shared-nothing store on raw NVMe wired into engines such as vLLM.

MinIO: you don’t need an object cache. You need a faster object store. MinIO blogs https://www.min.io/blog/you-dont-need-a-cache-you-need-a-faster-object-store about object caching in AI and says you don’t need a cache if your object storage is fast enough. Blogger Daniel Valdivia, a MinIO architect, compares a CoreWeave LOTA https://www.blocksandfiles.com/ai-ml/2025/10/16/coreweaves-lota-tech-moves-object-data-at-high-speed/1597613 Local Object Transport Accelerator caching proxy benchmark with MinIO AIStor https://www.blocksandfiles.com/ai-ml/2025/03/17/minio-aistor-20-embraces-nvidia-gpudirect-bluefield-3-and-nim-for-faster-ai-inference/1590187? gl=1 1st44id ga MTY2OTcyMjAyNS4xNzcwODg2MTMy ga NSDTXHMMN0 czE3ODUzMTY2MzckbzE4NiRnMSR0MTc4NTMyMDEzOSRqNTQkbDAkaDA. object storage performance. CoreWeave ran Warp https://github.com/minio/warp , an open benchmark test, loading object data on on 20 GPU nodes, each with 8 GPUs, 1 TiB of LOTA cache per node, and a dedicated network adapter with dual 100 Gbps links for storage access, and achieved S3 GET performance of 18.4 GiB/s per node 19.8 GB/s . MiniO’s AIStor, using Solidigm QLC SSDS, with no caching achieved c33.5 GiB/s per node 36 GB/s in a separate benchmark https://www.solidigm.com/content/dam/solidigm/en/site/products/technology/2026/new-economics-of-ai-factory-minio/document/new-economics-of-ai-factory-graph.pdf . Valdiva says: “An erasure-coded object store on capacity-optimized QLC drives, running over plain TCP, delivers more throughput per node than a warm cache on local NVMe about 33.5 GiB/s versus 18.4 GiB/s , and more than 11 times the cluster throughput of CoreWeave's own cold path. A cache should be faster than what's behind it. When a durable store with parity beats the cache, the architecture is telling you something.” He explains: “A cache exists because the thing behind it is slow. When your object store is fast, you don't need to make complicated trade-offs to keep GPUs fed. …you shouldn't need a cache on every GPU node to make your object store fast in the first place.” MinIO is not saying you don't need KV caching systems: “An object cache speeds up S3 GETs of datasets and checkpoints. An inference KV cache stores attention state so the engine doesn't recompute a long prompt every time a user comes back. They have different access patterns, block sizes, and success metrics.” They are two different caching situations. MinIO supports KV caches and has developed its own MemKV, https://www.blocksandfiles.com/ai-ml/2026/05/12/minio-adds-petabyte-scale-memkv-cache-for-nvidia-gpu-inference/5238593 a shared-nothing KV store on raw NVMe, wired into inference engines like vLLM. In effect, CoreWeave wouldn't need LOTA if it used MinIO's AIStor instead.