WEKA NeuralMesh GPU Storage 2026: UK Buyer Guide

What Andromeda and WEKA have actually done

Andromeda operates as a broker and manager of GPU capacity, routing training and inference workloads across more than 50 compute providers and, by its own account, serving over 100 AI customers across billions of GPU-hours. Every cluster it delivers is certified against a common benchmark covering GPU, storage, network fabric and security before it reaches a customer.

The problem WEKA's NeuralMesh was brought in to fix is a familiar one for anyone buying capacity across multiple GPU clouds: each provider historically ran its own storage stack, and performance varied cluster to cluster even when the GPUs themselves were identical. NeuralMesh replaces that patchwork with a consistently configured data layer, deployed either as a dedicated NeuralMesh cluster or as a GPU-native configuration called NeuralMesh Axon.

How NeuralMesh Axon turns spare NVMe into shared storage

Axon is the part of this story most relevant to procurement teams. Rather than requiring separate storage appliances, it converts the local NVMe drives, CPU cores, DRAM and NICs already inside a GPU server into a unified high-performance data pool. Roughly half of Andromeda's NeuralMesh deployments now run this Axon configuration, with the remainder reserved as dedicated deployments for customers who need exclusive GPU compute and memory.

The economic logic is straightforward: the hardware is already racked, powered and paid for, so extracting a shared storage layer from it adds performance without adding footprint or extra cost. WEKA cites a research lab whose workloads were bottlenecked by a cluster's native storage; Andromeda resolved it by deploying Axon on the same hardware rather than provisioning anything new. Deployment via the NeuralMesh Kubernetes Operator reportedly takes about 10 minutes from bare metal, which Andromeda's lead architect, Vishvajit Kher, described as the template for how new clusters now come online.

The numbers UK buyers should sanity-check

Andromeda reports NeuralMesh Axon clusters sustaining more than 400 GB/s of aggregate throughput and exceeding 6 million IOPS in production. In a biotech workload involving roughly one billion files per directory, data transfer times reportedly fell from hours to minutes, and environment startup times dropped from around three minutes to 30 seconds. For teams running iterative training and inference cycles, faster startup translates directly into more experiments per development window — a metric worth modelling against your own workload cadence using tools like the AI GPU calculator before assuming the same uplift applies to your estate.

Standardisation cuts both ways: the lock-in question

The case for standardising storage across a multi-provider GPU fleet is easy to make: predictable IOPS and throughput regardless of which hyperscaler or neo-cloud is underneath. But UK buyers evaluating self-hosting LLM vs cloud GPU economics should read this the other way too. Once a broker or provider standardises its entire fleet on one vendor's data layer, migrating workloads or data out of that fleet means migrating out of NeuralMesh as well — and WEKA's own roadmap keeps moving, from general availability of its enterprise NeuralMesh AI Data Platform in March 2026, through BlueField-4 DPU integration, to NeuralMesh 6 and an exabyte-scale WEKAPod array announced for general availability in the second half of 2026.

That pace of change is a feature if you're on the upgrade path with no additional cost, as existing customers reportedly are. It's a risk if your contract, data egress terms, or sovereignty requirements assume a static architecture. Buyers should ask any GPU provider standardising on a single storage vendor exactly how data portability and format compatibility are guaranteed across future NeuralMesh versions.

Where this lands for UK procurement

WEKA already has UK-relevant traction: NexGen Cloud signed on as a GPU-as-a-service customer in 2024, alongside earlier deals with Applied Digital and Iris Energy, well before the Andromeda integration. That history suggests the Andromeda deployment is an extension of an established GPU-cloud playbook rather than a first foray. Separately, sovereign and regulated deployment models — including dedicated sovereign, regulated multi-tenant, and hybrid sovereign configurations reported in other markets — show NeuralMesh being positioned for regimes with data residency constraints, which is directly relevant to any organisation weighing sovereign AI build or rent decisions.

For UK buyers running procurement through IT procurement services, the practical takeaway is to treat storage standardisation claims as a checklist item, not a marketing line: ask for the certified benchmark methodology, the fault-isolation guarantees for multi-tenant clusters, and whether object-tier backends (WEKA has validated Scality RING for this role) are available to control long-term storage cost as data volumes grow. Buyers still comparing shared-storage architectures more broadly can use a storage solution finder alongside vendor-specific due diligence, and revisit AI and Analytics Storage options if workloads are metadata-heavy rather than pure throughput-bound.