GPU memory is the most expensive resource in production AI and the one running out fastest. Long context windows and multi-turn conversations force AI models to repeatedly recompute information they have already processed, consuming GPU memory and compute that could otherwise serve additional users or generate new responses. Instead of treating GPU memory as the limiting factor, why not extend it with much cheaper storage technologies? Weka, a company specializing in data infrastructure for AI, believes that low-cost NAND flash can close that gap. Its new NeuralMesh 6 software platform, launched alongside the first self-designed hardware line called Wekapod 3, extends what Weka calls Augmented Memory Grid, an approach that aggregates NAND flash to behave like GPU memory at a fraction of the cost.
NeuralMesh 6 and Wekapod 3: Weka's New Hardware and Software for AI Infrastructure
NeuralMesh 6 introduces four capabilities aimed at a functionality gap that CEO and co-founder Liran Zvibel says has been costing Weka deals in competitive evaluations. First, composable and virtual multi-tenancy. Composable clusters give anchor tenants full hardware-level isolation with dedicated CPU, memory, and storage. Virtual multi-tenancy runs over Weka's RDMA fabric, delivering network-level isolation that scales past 1,000 tenants per cluster, with provisioning in under 30 minutes. Combined, a single cluster running 50 composable clusters can support up to 50,000 tenants. Second, unified file and object storage. Most storage systems maintain two separate paths: a file-based path (used heavily in training and fine-tuning) and an object-based path (S3 format expected by inference and cloud-native tools). Weka claims the same physical data on disk is directly readable through either path at once, no translation layer, no second copy. Zvibel targets non-AWS GPU clouds such as Lambda, Nebius, G42, and CoreWeave, promising roughly two orders of magnitude higher performance than conventional S3 and a capacity-based pricing model instead of per-API charges. Third, metadata-first replication: destination environments become browsable before a full data copy arrives, with data hydrating only when accessed. Fourth, AlloyFlash mixes TLC and QLC NAND flash within a single cluster, automatically routing latency-sensitive work to TLC while running bulk-capacity workloads on QLC, cutting cost per terabyte without performance penalty. Data reduction now runs by default rather than as an option.
Sponsored Protocol
Augmented Memory Grid: Caching 100% of Pre-Calculated Tokens
Multi-tenancy and object storage solve how enterprises and neo clouds operate the platform day to day. A harder problem sits underneath: as context windows and multi-turn interactions grow, so does GPU compute wasted recalculating work a model has already done. Augmented Memory Grid, a NeuralMesh 6 feature built specifically for this, is Weka's answer. Every prompt triggers two stages: prefill, which calculates attention (the core mechanism behind how large language models process input) and is computationally expensive, and decode, which converts that calculation into output and is comparatively lightweight. The cost shows up hardest in multi-turn sessions like chat or coding, where each new turn re-triggers prefill for everything that came before it, unless that work has been cached. Zvibel explains: "If you have 10 turns, you may overcalculate 100 times because you are redoing all of them. If you have 20, you will overcalculate 400 times. You can put two orders of magnitude more NAND than you could afford in shared memory, and we can cache 100% of the pre-calculated tokens, so you never need to redo it." This translates into better utilization of existing GPU investments, lower inference costs, and faster deployment of new AI workloads without waiting months for additional GPU capacity.
Sponsored Protocol
Competitive Positioning and Data Reduction Guarantees
The storage industry is rapidly repositioning around AI. Dell, NetApp, Pure Storage, and VAST have all shifted toward AI infrastructure. Steve McDowell, chief analyst at NAND Research, notes: "The storage world is shifting its focus from serving bits to enterprise workloads to managing data at the speed of AI. Companies like Weka and VAST are the true AI-native data companies, solving these problems since day one." McDowell highlights Augmented Memory Grid as Weka's clearest technical lead: "Weka continues to have the most technically capable KV cache implementation on the market with its Augmented Memory Grid. They were early with this technology and continue to innovate. This is critical for AI inference, as it enables a level of GPU efficiency that, without question, saves money on GPUs and memory." He also flags Weka's contractual guarantee on its data reduction claims as underappreciated: "Weka is putting its money where its mouth is with its contractual guarantees for its data reduction promises." For buyers, the advice is to look hard at what vendors promise versus what they deliver, by talking to organizations running similar workloads at similar scale. Weka targets organizations already operating AI at scale or expecting rapid growth, particularly enterprises building internal copilots, customer service agents, software engineering assistants, or retrieval systems with long context windows. For those with GPU utilization already a limiting factor, the benefits are immediate.
Sponsored Protocol
For more on AI security, read our article on OpenAI's autonomous AI hacker. For technical details on flash memory, see the Wikipedia page on NAND flash.
Sponsored Protocol
