English
WEKA has unveiled two interconnected products: NeuralMesh 6, its most significant software release to date, and WEKApod 3, a new generation of storage appliances purpose-built to run the platform. NeuralMesh 6 can be deployed on customer hardware, while WEKApod serves as a turnkey solution shipped with preinstalled software.
NeuralMesh 6 delivers native multi-tenancy, unified file-and-object protocol stack, metadata-driven data mobility, always-on data reduction with contractual guarantees, Kubernetes-native operations and integrated observability. Custom-engineered WEKApod 3 delivers industry-leading capacity and performance density within a single rack and comes in three variants: Nitro for peak performance, Prime for balanced capacity, and Prime Max for maximum density.
WEKA NeuralMesh 6
Both launches reflect a key market shift. As AI workloads shift from training to production-scale long-context, agentic and retrieval-based inference, storage and memory infrastructure — rather than GPU count alone — increasingly determine cost per token and system throughput.
NeuralMesh 6: Multi-Tenancy, Unified Protocols and Data Mobility
NeuralMesh 6 consolidates capabilities AI operators previously sourced from multiple vendors into a single software stack: multi-tenancy, unified file and object protocols, cross-site data mobility, continuous data reduction, Kubernetes-native management and embedded observability.
Its multi-tenancy consists of two combinable layers. Composable Clusters offer hardware-level isolation with dedicated CPU, memory and drives for tenants requiring resource guarantees and stable performance. Virtual Multi-Tenancy provides VPC-style network isolation via WEKA’s Virtualized RDMA Data Fabric, supporting private VLANs, overlapping IP ranges, per-tenant QoS, isolated encryption key management and independent LDAP or Active Directory authentication. A single cluster supports over 1,000 isolated logical tenants, with new tenants provisioned within 30 minutes. Combined, one WEKA cluster hosting 50 Composable Clusters can accommodate up to 50,000 isolated tenants on shared physical infrastructure, enabling seamless scaling without system redesign.
On the protocol front, NeuralMesh 6 features native S3 support, allowing identical data blocks to be accessed simultaneously via S3 and POSIX without protocol gateways. Files written through NFS/POSIX are immediately accessible via S3 and vice versa, removing redundant dataset copies generated during training, fine-tuning and inference. Optimized for AI access patterns, each node supports 2,000–5,000 concurrent S3 connections — roughly five times that of traditional S3 architectures. S3 over RDMA enables zero-copy transfers directly into GPU memory.
Metadata-first replication powers data mobility. Target environments become browsable instantly instead of waiting for full data migration; data loads on demand to cut WAN traffic, letting enterprises deploy workloads wherever GPU resources are available. This release adds asynchronous replication and remote caching, laying groundwork for cross-site and cross-cloud federation alongside a global namespace.
The replication technology already runs in production. WhiteFiber CEO Sam Tabar explained that NeuralMesh’s intelligent replication makes datasets accessible across locations and delivers required data to newly assigned GPUs, the same architecture underpinning Project Redwood, the 111.2 Tbps inter-datacenter supercluster covered recently.
Up to 6X Capacity Savings Through Data Reduction
NeuralMesh 6 enables default data reduction including fingerprinting, similarity hashing, deduplication and compression across all deployments. It carries write overhead below 5%, delivers up to sixfold capacity savings on AI training data, and includes contractual guarantees for reduction ratios and performance impact. A new Kubernetes Operator automates cluster deployment and lifecycle management, shortening rollout from weeks to hours. NeuralMesh Observe, provided at no extra cost, offers SaaS multi-cluster dashboards, client diagnostics and alerts for Slack, PagerDuty or email.
WEKA has deployed its Augmented Memory Grid in production. It effectively expands GPU memory by accelerating persistent KV cache access to NeuralMesh-managed NVMe storage on Oracle Cloud Infrastructure. Benchmarks on OCI H100 hardware recorded 10× higher token throughput, 10× more concurrent users and 7× more tokens per GPU compared with DRAM-only alternatives. Pablo Selem, OCI senior director of software development, noted the technology eliminates memory bottlenecks to boost throughput and user capacity on existing GPU hardware.
WEKApod 3: Custom Hardware Optimized for the Software
WEKApod 3 is WEKA’s proprietary hardware platform, not a reference design built on third-party OEM chassis. The vendor states one WEKApod rack delivers 1.1 exabytes effective capacity from 441.5 PB raw storage, making it the first single-rack system to surpass one exabyte effective capacity, with rack throughput reaching 10.2 TB/s and 210 million IOPS. WEKA claims it achieves 267% higher effective capacity density and 114% greater throughput density per rack unit than competing publicly available systems.
The design adopts PCIe Gen 6 internal fabric, cable-based drive interconnects instead of backplanes, NVIDIA ConnectX SuperNICs for Spectrum-X Ethernet, and software-controlled thermal systems rated for 35°C ambient temperature. Under thermal stress, it throttles NVMe power rather than triggering shutdown. Multiple patents are pending for its chassis, interconnect, thermal management and serviceability design. Hot-pluggable boot drives support GUI-guided replacement completed in roughly 10 minutes instead of lengthy maintenance windows, alongside headless, cloud-driven rack-scale deployment via NeuralMesh Home.
Three Configurations for Different Workload Priorities
WEKApod Nitro targets bandwidth-critical workloads to keep GPUs fully utilized. It uses a 2U four-node chassis with four independent failure domains and 56 TLC drives, paired with dual-port NVIDIA ConnectX networking delivering 800 Gb/s throughput. WEKApod Prime balances capacity and performance with AlloyFlash mixing TLC and QLC drives within a similar 2U four-node chassis supporting 56 drives. WEKApod Prime Max prioritizes compact high density: a 2U two-node chassis holding 70 Micron 245.76 TB 6600 ION NVMe SSDs. Combined with NeuralMesh object storage and data reduction, it reaches 1.1 exabytes effective capacity within a single 56U rack.
AlloyFlash enables Prime and Prime Max configurations. It automatically directs latency-sensitive tasks to TLC flash and bulk data to lower-cost QLC media, around 30–40% cheaper per TB, without manual tuning. It highlights tight synergy between NeuralMesh 6 and WEKApod 3: the software’s tiering intelligence allows high-density hardware to deliver production-grade performance beyond theoretical capacity metrics.
WEKA’s in-house hardware strategy responds to widespread datacenter constraints. U.S. datacenter construction fell in 2025 for the first time since 2020; grid connection waiting periods last four to seven years in major markets, and Morgan Stanley forecasts a 49-gigawatt U.S. power shortage through 2028. These issues are compounded by ongoing NAND supply limits and extended OEM lead times. WEKA contends inefficient storage competes with GPUs for limited rack space and power. By controlling its hardware supply chain instead of relying fully on OEM channels, it aims to offer more predictable pricing and delivery timelines for large-scale infrastructure projects.
Jason Hardy, NVIDIA VP of Storage Technology, said Spectrum-X Ethernet provides WEKApod 3 with the high-bandwidth, low-latency fabric required for stable large-scale storage-to-GPU data transmission. Steve McDowell, chief analyst at NAND Research, pointed out production inference presents different infrastructure challenges from training. Key metrics now include tokens per rack, tokens per watt and sustained cost per inference, which buyers should use to evaluate vendors. Jeremy Werner, Micron Core Data Center Business Unit SVP and GM, added that the new WEKApod platform with Micron’s 245TB SSDs fits 15.8 petabytes into a 2U enclosure, saving power and rack space for additional compute resources.
Availability
NeuralMesh 6 is scheduled for general availability in H2 2026. Existing WEKA customers can upgrade free of charge via standard channels. WEKApod Nitro, Prime and Prime Max are open for orders through WEKA distributors and VARs, with shipments starting in fall 2026 and NeuralMesh 6 preloaded. This generation introduces configurable WEKApod SKUs for the first time, letting customers select chassis, memory, drive capacity and quantity to build deployments ranging from under 1PB to 100PB and above.
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!
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!