WEKA has launched the exabyte-scale WEKAPod all-flash array, optimized for its sixth-generation NeuralMesh software. The new platform delivers massive data throughput to power production-grade AI training and inference workloads. It positions WEKA alongside premium AI storage vendors including DDN, Everpure, NetApp and VAST Data, featuring industry-leading density of 1.1 exabytes of effective capacity within a single 56 RU rack.
NeuralMesh v6 integrates native hyperscale multi-tenancy, an NVMe-based unified file and S3 object stack, intelligent data mobility with replication and remote caching, performance-guaranteed always-on data reduction, Kubernetes-native workflows, and unified observability. On Oracle Cloud Infrastructure (OCI) H100 hardware, its Augmented Memory Grid delivers 10x higher token throughput, 10x more concurrent users, and 7x higher per-GPU token output than NVIDIA CMX in production. For CoreWeave deployments, it boosts per-GPU token throughput by 4.2x and cuts TTFT latency by up to 6x versus CMX.
WEKA co-founder and CEO Liran Zvibel commented in July 2026: “Legacy AI infrastructure built for traditional workloads cannot support modern inference-era operations. Today’s AI profitability hinges on rack space efficiency, energy consumption, supply chain stability and operational density.
“WEKAPod is purpose-built to fit these constraints. Engineered exclusively for inference-era computing with a self-controlled supply chain, it delivers upfront cost advantages and superior TCO across full-stack AI infrastructure needs, setting a new benchmark for large-scale production AI infrastructure evaluation.”
The WEKAPod is a dedicated storage appliance that feeds data to GPU clusters such as NVIDIA SuperPODs, integrating pre-configured hardware and WEKA’s data platform software. First launched in late 2025 with PCIe Gen4 Prime and Gen5 Nitro models featuring AlloyFlash tiered SSDs, the updated third-generation lineup adds 2RU-based Nitro, Prime and new Prime Max configurations tailored for production AI inference.
Prime Max is WEKA’s highest-capacity model, exceeding 1 EB per rack. Prime delivers balanced price-performance via four 2U rack servers with 70 mixed QLC/TLC NVMe drives managed by AlloyFlash. Nitro targets high-performance AI factories and GPU-as-a-service environments with hundreds to thousands of active GPUs.
The third-gen WEKAPod is the industry’s first single-rack system to surpass exabyte-scale effective capacity. It packs 441.5 PB of raw NVMe storage to deliver 1.1 EB effective capacity via NeuralMesh data reduction, alongside 10.2 TB/s per-rack throughput and 210 million rack-level IOPS.
Its upgraded hardware stack includes PCIe Gen6 internal fabric, backplane-free cable-based drive interconnects, NVIDIA ConnectX SuperNICs with Spectrum-X Ethernet connectivity, and software-defined thermal management. The platform minimizes data center footprint, optimizes rack space utilization and improves tokens-per-watt efficiency without compromising throughput or workload performance. WEKA states it outperforms competing solutions by 267% in effective capacity density and 114% in throughput density per rack unit.
By directly managing component sourcing without OEM channel reliance, WEKA provides AI cloud providers and enterprises with stable pricing and predictable lead times for large-scale infrastructure deployments. All WEKAPod units ship factory-configured and pre-validated with full NeuralMesh v6 capabilities including Augmented Memory Grid, eliminating the need for dedicated on-site storage teams.
NeuralMesh v6, WEKA’s parallel file system, leverages Augmented Memory Grid to extend GPU memory for inference workloads. It externalizes storage as a low-latency (microsecond-level) KV cache with multi-TB/s bandwidth, delivering petabyte-scale extended address space for AI models.
Key NeuralMesh 6 capabilities include:
Composable Clusters: Delivers full hardware isolation with dedicated CPU, memory and storage for anchor tenants, ensuring guaranteed resource allocation, strict workload separation and consistent performance under heavy loads.
Virtual Multi-Tenancy: Offers VPC-grade isolation via Virtualized RDMA Data Fabric (VRDF), supporting private VLANs, overlapping IP spaces, per-tenant QoS, independent KMS encryption and LDAP/AD authentication. It scales to over 1,000 isolated logical tenants per cluster with new tenant provisioning under 30 minutes.
Shared physical NVMe drives across multiple NeuralMesh clusters maximize device parallelism for multi-tenant inference. A single hardware cluster supporting 50 Composable Clusters can host up to 50,000 isolated logical tenants. It unifies S3 and POSIX file access under one namespace, enabling zero-copy S3-over-RDMA data transfers directly to GPU memory with 2,000–5,000 concurrent node connections — roughly five times conventional S3 concurrency.
Its AlloyFlash technology transparently unifies cost-effective QLC and high-performance TLC NVMe flash in one cluster, routing latency-sensitive workloads to TLC and bulk capacity tasks to QLC at 30–40% lower per-TB cost.
Always-on data reduction combines fingerprinting, similarity hashing, deduplication and compression with under 5% write overhead, delivering up to 6x capacity savings for AI training data. WEKA provides contractual guarantees for both reduction ratios and performance impact, with pre-sales workload analysis to forecast real-world efficiency.
All data reduction processes run off the critical write path via background similarity compression and cross-filesystem deduplication, enabling native NVMe-speed writes regardless of capacity usage or background load. This avoids the write performance degradation common to inline reduction architectures during checkpoint-heavy AI training, mirroring ExaGrid’s deduplication-free landing zone design for consistent performance.
NeuralMesh v6 supports full federation and global namespace capabilities. Its asynchronous replication enables AI data mobility, cloudbursting and multi-site collaboration, with priority metadata replication for instant destination environment browsing and on-demand data hydration. This reduces WAN bandwidth consumption via replicated compressed data.
A dedicated Kubernetes Operator automates cluster deployment and lifecycle management for Kubernetes-based AI clouds and research labs. The SaaS-based unified Observe platform delivers multi-cluster dashboards, client-level diagnostics, configurable intelligent alerting, and integrated routing to Slack, PagerDuty or email with direct dashboard links.
WEKA Chief Product Officer Ajay Singh stated: “Today’s production AI operators rely on disjointed, retrofitted third-party stacks with siloed file/object layers, manual data movement and bolted-on multi-tenancy. NeuralMesh 6 solves these pain points with a unified platform that converges high-performance file and high-capacity object storage on shared hardware, with native multi-tenancy, intelligent data mobility and built-in data efficiency. It is purpose-built for production inference workloads, not retrofitted for them.”
NeuralMesh 6 will reach general availability in H2 2026, with free seamless upgrades available to all existing WEKA customers via standard update channels.
Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!
Your message must be between 20-3,000 characters!
English
