WEKA announces NeuralMesh 6 for production AI infrastructure

The release combines several capabilities that infrastructure teams often operate as separate systems:

• •Native physical and logical multi-tenancy, including per-tenant networking, encryption and authentication
• •File and S3 access to the same physical data blocks on NVMe storage
• •Metadata-first asynchronous replication and remote caching across sites
• •Always-on deduplication, compression and related data-reduction features
• •A Kubernetes operator for deployment and lifecycle management
• •SaaS-based multi-cluster observability through NeuralMesh Observe

WEKA says virtual multi-tenancy can support more than 1,000 logical tenants per cluster, while combining it with hardware-isolated composable clusters can raise the total to 50,000. It also claims less than 5% write overhead and up to 6x capacity savings for its always-on data reduction. Those figures are vendor specifications rather than results independently reproduced in the retrieved reporting.

The release's central inference claim concerns WEKA's Augmented Memory Grid, which stores persistent attention KV cache on NeuralMesh-managed NVMe and accelerates access to it. Reusing cached state can reduce repeated prefill work for long-context or shared-prefix requests, potentially improving GPU utilization.

WEKA reports that benchmarks on Oracle Cloud Infrastructure H100 systems produced 10x higher token throughput, 10x more concurrent users and 7x more tokens per GPU than DRAM-based alternatives. StorageReview and SiliconANGLE reported the same figures, but the retrieved material does not disclose the workload mix, model, request lengths, cache-hit rate, baseline sizing, latency percentiles or cost assumptions needed to reproduce the comparison. The numbers should therefore be treated as vendor-reported performance claims.

The architectural direction is relevant: inference systems increasingly need to coordinate model data, retrieval stores, session state and persistent cache across GPUs and locations. A buyer should still test NeuralMesh 6 against its own serving pattern. Useful measurements include time to first token, decode latency, tail latency, throughput per GPU, cache-hit rate, recovery behavior and tenant isolation under noisy-neighbor load.

Teams should also validate interoperability with their inference server, Kubernetes environment, identity system and monitoring stack. The release may reduce the number of storage and operations components required, but its economic value depends on workload-specific performance, licensing and recovery characteristics that WEKA has not fully detailed in the public announcement.