WEKA Launches NeuralMesh 6 and WEKApod 3 as AI Infrastructure Shifts Toward Inference

WEKA has introduced a coordinated software and hardware overhaul aimed at one of the AI industry’s most expensive emerging problems: keeping GPUs productive as models move from training into continuous, large-scale inference.

The announcements include NeuralMesh 6, a major update to the company’s data and memory platform, alongside third-generation WEKApod appliances engineered for AI clouds, model developers, research organizations, and enterprises operating GPU infrastructure.

Together, the releases reflect a broader change in AI infrastructure design, with storage evolving from a passive repository into an active memory and data-delivery layer.

A Unified Push Into Production AI

The infrastructure requirements of AI inference differ considerably from those of model training. Training jobs typically process large datasets through relatively predictable pipelines. Agentic AI, retrieval-augmented generation, and long-context reasoning systems must instead access changing data, maintain context across repeated interactions, and support many concurrent users without leaving costly GPUs waiting for information.

WEKA’s response is to treat storage, memory extension, file access, object access, and data movement as parts of the same platform. NeuralMesh is designed to provide a shared data foundation for training, inference, and high-performance computing across on-premises environments, public clouds, and hybrid deployments.

The new release expands that architecture with multi-tenancy, native object storage, distributed caching, automated data reduction, Kubernetes operations, and centralized observability.

Rather than positioning the software and appliance announcements as unrelated upgrades, the company is presenting them as two parts of the same system. NeuralMesh 6 controls how data is stored, moved, isolated, and delivered, while WEKApod 3 provides hardware designed around those functions.

NeuralMesh 6 Unifies File and Object Workloads

One of the more consequential additions in NeuralMesh 6 is a native S3 implementation running on NVMe storage. The same underlying data blocks can be accessed through S3 object protocols and POSIX or Network File System interfaces without maintaining separate copies.

That matters because AI pipelines often move information between object storage, high-performance file systems, training environments, and inference clusters. Every copy consumes capacity, introduces delays, and complicates governance. A unified namespace could allow data created through one protocol to become immediately available through another.

WEKA says its implementation supports between 2,000 and 5,000 concurrent S3 connections per node. It also supports S3 over Remote Direct Memory Access, allowing data to move toward GPU memory without following the conventional path through multiple CPU and operating-system layers.

The platform introduces two levels of multi-tenancy. Composable Clusters allocate dedicated processor, memory, and storage resources to individual tenants, while virtual multi-tenancy provides network isolation, independent encryption, separate authentication, and per-tenant quality-of-service controls.

Virtual environments can support more than 1,000 isolated tenants per cluster, according to the company. Combining the physical and virtual models could allow a large infrastructure operator to support tens of thousands of logically separated customers without deploying an independent storage cluster for each one.

This is particularly relevant for GPU-as-a-service providers and sovereign AI clouds that need to balance high infrastructure utilization with strict customer isolation.

Moving Data to Wherever GPUs Are Available

NeuralMesh 6 also introduces metadata-first replication and remote caching for AI workloads distributed across data centers, clouds, and geographic regions.

Traditional replication generally requires an entire dataset to be copied before a workload can begin. NeuralMesh instead makes the destination’s metadata immediately visible, then transfers the underlying data as it is requested.

This could let operators move jobs toward available GPU capacity without first replicating every file associated with the project. The approach is intended to support cloud bursting, distributed AI infrastructure, and collaboration between geographically separated research or engineering teams.

It also represents an early step toward a global namespace model in which data can be addressed consistently regardless of where it was originally stored.

Another new feature applies fingerprinting, similarity hashing, deduplication, and compression continuously rather than treating data reduction as an optional background process.

WEKA claims NeuralMesh 6 can provide up to sixfold capacity savings on AI training data with less than 5% write overhead. Actual reductions will depend on the dataset. Checkpoints, container layers, model versions, and iterative training data can contain significant repetition, while other data types may compress less effectively.

The company plans to analyze representative customer data before deployment and use the resulting estimate as part of its capacity and performance commitments.

Turning Storage Into an Extended AI Memory Layer

NeuralMesh also incorporates WEKA’s Augmented Memory Grid, which uses NVMe storage as a persistent extension of GPU memory for inference.

Large language models create a key-value cache containing information calculated from a prompt or conversation. For long contexts and agentic workflows, that cache can become extremely large. When it cannot remain in high-bandwidth GPU memory, systems may need to discard it, recompute it, or move it into slower infrastructure.

Augmented Memory Grid stores that cache in a persistent NVMe-backed layer and uses Remote Direct Memory Access and NVIDIA GPUDirect Storage to move it back to GPUs.

The objective is to preserve reusable context while reducing repeated prefill computation, one of the more resource-intensive stages of long-context inference.

WEKA reports that testing on Oracle Cloud Infrastructure using NVIDIA H100 GPUs produced ten times higher token throughput, ten times more concurrent users, and seven times more tokens per GPU.

Those figures are workload-specific benchmarks rather than universal performance expectations, but they illustrate why persistent cache management is becoming an important part of inference infrastructure design.

WEKApod 3 Takes Control of the Hardware Layer

While previous WEKApod systems combined WEKA software with prevalidated infrastructure, the third generation moves further into vertically integrated system design.

WEKA designed the new two-rack-unit chassis, drive interconnect, cooling architecture, failure domains, and serviceability system. The appliances use PCI Express Gen 6 internally and NVIDIA ConnectX networking for connection to Spectrum-X Ethernet infrastructure.

The WEKApod family now consists of three configurable systems. Nitro is optimized for bandwidth-intensive inference and AI factory workloads. Prime combines triple-level cell and quad-level cell flash to balance performance and capacity. Prime Max prioritizes maximum storage density, fitting as many as 70 NVMe drives into a two-rack-unit chassis.

At full rack scale, WEKA says Prime Max can provide 441.5 petabytes of raw capacity and 1.1 exabytes of effective capacity after NeuralMesh data reduction. The rack-level system is rated for up to 10.2 terabytes per second of throughput and 210 million input/output operations per second.

The distinction between raw and effective capacity is important. The exabyte figure depends on the reduction achievable across the stored data and should not be interpreted as more than an exabyte of physical flash.

Even so, greater density can have meaningful consequences in facilities where storage competes with GPUs for rack space, cooling, and electrical capacity.

Designed for Constrained Data Centers

The new systems also address the physical limitations increasingly shaping AI infrastructure projects.

GPU clusters require large amounts of electricity, cooling, networking, and floor space. Storage systems that consume these resources inefficiently can reduce the number of accelerators that a facility can support.

WEKA claims the new appliances provide 267% greater effective capacity density and 114% greater performance density than the next-best publicly available alternatives in their respective categories.

The company has also designed the systems to operate in ambient temperatures of up to 35 degrees Celsius. Software-controlled power throttling is intended to reduce performance gradually during thermal stress rather than triggering a shutdown.

A backplane-free drive design creates smaller failure domains, while hot-pluggable boot drives and guided replacement procedures are intended to reduce maintenance time. Customers can configure chassis type, memory, drive capacity, and drive count, with deployments ranging from less than one petabyte to more than 100 petabytes.

Building an Ecosystem Around AI Factories

The accompanying partner announcements suggest the platform will reach customers through infrastructure integrators, cloud providers, and AI orchestration vendors.

Spectro Cloud is positioning NeuralMesh as a data layer that can be combined with PaletteAI for deploying and operating enterprise AI factories. World Wide Technology and Computacenter plan to incorporate the systems into broader infrastructure projects, while Glocomp highlighted customer demand for better AI performance and more predictable infrastructure costs.

This ecosystem strategy is significant because most organizations are not assembling production AI environments from a single supplier.

A typical deployment may include GPUs, networking, storage, Kubernetes, model-serving software, observability, security, and governance products from multiple vendors. Jointly validated configurations can reduce integration work, although customers will still need to test performance using their own models, datasets, networks, and concurrency requirements.

What This Means for AI Infrastructure

The most important aspect of the announcement is not any individual capacity or throughput number. It is the growing convergence of storage, distributed caching, memory management, data mobility, and tenant isolation.

As AI agents retain longer histories, access more enterprise information, and operate continuously, the infrastructure surrounding GPUs will increasingly determine the cost of each useful response.

Adding more accelerators will not solve bottlenecks caused by repeated context processing, slow data movement, fragmented protocols, or underused storage capacity. The next phase of AI infrastructure will therefore depend as much on feeding and managing GPUs efficiently as it does on increasing raw compute.

NeuralMesh 6 is scheduled to become generally available during the second half of 2026, with existing customers eligible to upgrade through the standard software channel. The new WEKApod Nitro, Prime, and Prime Max systems are available to order, with deliveries expected to begin in fall 2026.