Weka launches NeuralMesh 6 to cut GPU load with flash-backed AI caching

Weka has released NeuralMesh 6, a new version of its storage software, alongside Wekapod 3, its first hardware line designed in-house. The company did not disclose pricing, but the launch targets a costly bottleneck in production AI: GPU memory consumed by repeated processing of long prompts and multi-turn sessions.

The pitch is that cheaper flash storage can act as an extension of GPU memory for some AI workloads. Weka calls the architecture Augmented Memory Grid, which aggregates NAND flash and uses it to cache pre-calculated tokens so GPUs do less redundant work during inference. Weka co-founder and CEO Liran Zvibel told VentureBeat the company can cache “100% of the pre-calculated tokens,” reducing the need to recompute prior context in conversations or coding sessions.

That claim lands in a crowded part of AI infrastructure. Dell, NetApp, Pure Storage, VAST and Weka have all pushed harder into AI systems as enterprises and GPU cloud operators try to raise utilization on scarce accelerators. Weka’s argument is that storage can reduce the pressure to add more GPUs, particularly for customers already running large-scale inference or expecting usage to climb.

What NeuralMesh 6 adds

NeuralMesh 6 introduces four main capabilities aimed at enterprises and GPU cloud operators. Weka says its composable clusters provide hardware-level separation for anchor tenants, with dedicated CPU, memory and storage. A second model, virtual multi-tenancy, runs over Weka’s RDMA fabric and is designed for network-level isolation across more than 1,000 tenants per cluster. According to the company, provisioning can be completed in under 30 minutes, and a cluster with 50 composable clusters can support as many as 50,000 tenants.

The release also combines file and object storage access to the same physical data. Weka says customers can read the same data through file interfaces used in training and fine-tuning workflows and through S3-style object interfaces used by inference and cloud-native tools, without a gateway or duplicate copy. Zvibel told VentureBeat the company is aiming this feature at GPU cloud providers outside AWS, including Lambda, Nebius, G42 and CoreWeave. He claimed performance is roughly two orders of magnitude higher than conventional S3 and said Weka uses capacity-based pricing rather than per-API charges.

Another feature, metadata-first replication, lets a destination environment become browsable before a full data transfer completes. Zvibel said data movement that previously could take days, weeks or in extreme cases a month can now let customers start using newly allocated GPUs within an hour, with data pulled in as needed.

Weka also added AlloyFlash and default data reduction. AlloyFlash mixes TLC and QLC NAND in one cluster, sending latency-sensitive workloads to faster TLC and bulk-capacity jobs to cheaper QLC. The company says this lowers cost per terabyte without hurting workloads that need speed. Its data reduction now runs by default.

The inference problem Weka is targeting

The technical bet centers on prefill, the expensive stage in which a large language model processes input before generating output. Decode, the output stage, is lighter. In multi-turn interactions, models may repeat prefill over prior context unless that work is cached.

Zvibel told VentureBeat that a 10-turn session can lead to 100 times as much repeated calculation and a 20-turn session to 400 times, because previous context keeps being processed again. Weka’s view is that NAND flash is cheap enough to hold far more cached context than shared GPU memory can economically support.

Competitive pressure

Steve McDowell, chief analyst at NAND Research, told VentureBeat that storage vendors have shifted from traditional enterprise workloads toward AI data requirements over the past 18 months, naming Dell, NetApp and Pure. He said Weka and VAST are more AI-native because they have been working on these problems from the start.

McDowell described Weka’s Augmented Memory Grid as the company’s clearest technical advantage and said its key-value cache implementation is among the most capable in the market. He also pointed to Weka’s contractual guarantees for data reduction claims as a notable part of the offering.

For buyers, the issue is proof. McDowell said enterprises should test vendor claims against deployments at similar scale and similar workloads. Weka has made a specific technical claim around token caching and GPU efficiency, but the company did not disclose customer adoption figures, pricing, revenue impact or independent benchmark results for NeuralMesh 6.

This story draws on original reporting from VentureBeat.