HPE Alletra Storage MP X10000 release 4 is generally available, lifting the supported limit from 8 nodes and JBOFs to 16 of each, taking maximum capacity from 11.8 PB to about 23 PB, and adding NFS file storage alongside object.
What was announced
- HPE Alletra MP X10000 release 4 is now generally available and raises the supported software limit from 8 nodes and 8 JBOFs to 16 of each.
- The release adds NFS file storage to the existing object capability, making the X10000 a unified unstructured data system. NFS v4.1 is supported.
- Maximum supported capacity rises from 11.8 PB to about 23 PB: each 2RU storage node can hold 24 x 61.44 TB drives, so 16 nodes give 23,592.96 TB.
- Release 4 adds RDMA support for file data transfers plus NVIDIA GPUDirect support, so the GPU KV cache is supported for faster data movement.
- Omdia testing of the RDMA accelerated X10000 for KV cache offload showed up to 20x faster time to first token and up to 17x higher effective inference throughput.
- Other additions include NVIDIA NIM multimodal support, an optional 2U GPU node with one NVIDIA L40S in the data path for metadata extraction and vector embeddings, 400 GbE switch capability, active/active bucket level replication between X10000 systems, KMIP external key management and TLS 1.3 enhancements.
- HPE also offers more flexible subscription terms, including 1, 6 and 7 year options.
- The cluster scale increase comes from a raised supported limit rather than from new processors or a basic software advance, and the block based Alletra MP B10000 added file storage before the X10000 did.
The ETON view
The interesting part of release 4 is not the capacity number, it is that the scale up is a support limit change on the same disaggregated architecture. Existing X10000 owners can double a cluster without changing the platform, which turns a would be forklift refresh into node and JBOF additions. That is exactly the kind of upgrade where the hardware line, not the licence, decides the total cost, and where buying nodes and 61.44 TB drives on the open market rather than a single bundled quote makes a visible difference.
The RDMA and GPUDirect work aims squarely at the real bottleneck in inference platforms, which is feeding GPUs rather than the GPUs themselves. Anyone running a RAG or agentic pipeline should measure time to first token before buying more accelerators, because KV cache offload to fast storage often buys more usable throughput per pound than another GPU node. Omdia's figures are vendor commissioned tests, so treat them as a reason to trial, not as a specification.
One caution on the optional 2U GPU node: a single L40S in the data path is a fine metadata and embedding engine, not an inference tier, so plan compute separately. We supply Alletra MP nodes, JBOFs, drives and L40S class GPUs new or certified refurbished, quote against lead time as well as price, and will value existing kit for buyback when a cluster grows rather than moves.
Related infrastructure
Category: Storage · Vendor: HPE, NVIDIA · Technology: Alletra Storage MP X10000, JBOF, NFS v4.1, RDMA, GPUDirect · Last verified: 18 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
