As AI shifts from model training to production inference, the challenge is no longer just building infrastructure, but operating it efficiently at scale. Running AI at production scale means solving two problems at once. The first is compute: GPUs, the fabric that connects them, and the Kubernetes platform that schedules, secures and governs everything running on them. The second is data: training corpora, embeddings, model weights, and the key-value cache behind every inference request.
The cost of getting it wrong is now well documented. Cast AI’s 2026 State of Kubernetes Optimization Report, drawn from telemetry across roughly 23,000 enterprise Kubernetes clusters, put average GPU utilization at just 5%. Some of that is overprovisioning. Much of it is GPUs waiting on the layers around them: scheduling, networking, and data pipelines that can’t keep accelerators fed.
That’s the backdrop for WEKA’s announcement of NeuralMesh 6 on July 21. We’ve been partnering with WEKA since March, and we think this release deserves attention for what it does to the data half of the production AI problem when paired with a Kubernetes-native orchestration layer.
What’s in NeuralMesh 6?
WEKA is calling NeuralMesh 6 the most significant software release in its history, and the feature list does indeed make for impressive reading. In a single software stack, it delivers native multi-tenancy at hyperscale, native S3 access running directly on NVMe alongside file services, metadata-first data mobility with asynchronous replication and remote caching, always-on data reduction backed by performance guarantees, Kubernetes-native operations, and unified observability across every deployment.
Until now, operators have had to assemble those capabilities from multiple products and vendors. NeuralMesh 6 brings these capabilities together in a single platform built for production inference and accelerated compute environments. WEKA frames the release around the hot topic of inference economics in the agentic era: as workloads shift from one-off training runs to always-on agents serving continuous inference, the data and memory layer determines whether your GPUs pay for themselves.
A year of momentum behind this release
All these features haven’t appeared quite out of nowhere.
NeuralMesh 6 lands a year after WEKA introduced NeuralMesh, and the intervening releases have tracked a clear direction of travel. In October 2025, WEKA announced a next-generation NeuralMesh architecture for NVIDIA BlueField-4, eliminating the need for standalone storage servers. At NVIDIA GTC in March, it took the NeuralMesh AI Data Platform to general availability and integrated its Augmented Memory Grid with the NVIDIA STX reference architecture, reporting 6.5x token production in the same GPU footprint. And just last month, WEKA published production benchmarks on Oracle Cloud Infrastructure showing 10x more concurrent users and 7x more tokens per GPU than DRAM-only configurations.
Our own partnership sits inside that arc. In March we announced a collaboration to simplify NVIDIA AI Data Platform deployments, made WEKA part of the expanded PaletteAI partner ecosystem, and published a joint reference architecture for RDMA-accelerated, AI-ready storage on Supermicro hardware. In other words, NeuralMesh 6 is the next step in a roadmap we’ve been building against for months.
The perfect pairing
GPUs only create value when they’re running models. Getting from racked, powered hardware to a governed, multi-tenant AI workload is an orchestration problem we’ve spent years solving.
PaletteAI is our declarative, Kubernetes-native platform for deploying and managing the full AI stack across data center, cloud and edge: the host operating system, Kubernetes itself, networking, GPU drivers and operators, model-serving frameworks, and the day-2 lifecycle management that keeps fleets consistent. With PaletteAI, a multi-layered stack like a WEKA-based NVIDIA AI Data Platform deploys in a single click, with the WEKA operator, CSI plug-in, secrets and configuration manifests applied identically on every cluster.
As AI environments scale, operators need more than orchestration. They need secure multi-tenancy, workload mobility, operational visibility, and efficient use of infrastructure across the entire stack. What makes the NeuralMesh 6 pairing work is where the two platforms meet: Kubernetes. NeuralMesh 6 ships with Kubernetes-native operations, and PaletteAI is built on a declarative Kubernetes model from the ground up. That shared foundation makes the data layer and the orchestration layer two halves of one repeatable, fleet-wide motion. We call that ‘rack to response’.
Mapping the capabilities to real outcomes
Several NeuralMesh 6 and PaletteAI joint capabilities line up with patterns we see across our partner ecosystem and customer base.
Multi-tenancy, for the people building clouds. Native multi-tenancy at hyperscale answers what neocloud and GPU-as-a-service providers need most: hard isolation between tenants without sacrificing utilization. Pair it with PaletteAI’s tenant cluster model and you get isolation that runs the full depth of the stack, from Kubernetes namespace to data path. That’s the difference between a demo and a sellable, defensible service.
Data mobility, for edge-to-core and sovereign AI. Metadata-first mobility with replication and remote caching suits organizations that need to train where the compute is and serve where the users (or the regulations) are. Train in a core data center, replicate intelligently, cache at the edge, and serve locally, all governed through one PaletteAI management plane spanning every location. For sovereign AI and public sector teams, that combination directly supports data residency requirements. Instead of moving petabytes of data upfront, NeuralMesh 6 synchronizes metadata first and retrieves data on demand. That makes it possible to move workloads quickly without waiting for full dataset replication.
Always-on data reduction, for GPU ROI. Data reduction with contractual performance guarantees means more effective capacity behind the same GPUs and the same power budget, without the latency penalty that usually makes teams switch reduction off. When fleets are averaging 5% utilization, every point of efficiency you claw back is money.
Unified observability, for day 2. NeuralMesh 6’s observability on the data side complements the fleet-wide operational visibility PaletteAI provides, so teams can reason about the whole system rather than chasing telemetry across silos.
The shorter path to production
The agentic era will be won by operators who can stand up production AI quickly, run it efficiently and govern it at scale. That requires more than orchestration alone; it requires shared infrastructure services spanning multi-tenancy, data efficiency, mobility, observability, and modern protocol support. NeuralMesh 6 takes serious complexity out of the data and memory layer. PaletteAI takes it out of orchestration and lifecycle. Together, that’s a shorter and more credible path from rack to response.
If you’re planning production AI infrastructure (building a neocloud, standing up sovereign capacity, or modernizing an enterprise platform), start with our partnership announcement, see how one-click AI Data Platform deployment works, or dig into the joint reference architecture on our design hub. And if you’d rather talk it through with someone who’s deployed this stack, get in touch.