WEKA

In this Vendor Showdown presentation moderated by Addison Snell (Intersect360 Research) and Rupak Biswas (NASA Ames Research Center), Shimon Ben David of WEKA presents the company's high-performance storage platform for AI workloads. He positions WEKA as shared distributed storage at petabyte to exabyte scale that outperforms local NVMe drives, citing mount points delivering hundreds of gigabytes per second and over three million IOPS, with standings in benchmarks such as IO500 and SpecSFS. The core theme is inference at scale: with trillions of tokens generated weekly industry-wide, the bottleneck is the memory wall, as GPUs exhaust HBM capacity for KV cache during prefill and decode and must recompute context. WEKA's Augmented Memory Grid (AMG) extends KV cache onto its storage layer, which the company claims is faster than local DRAM for prefill operations. Cited customer results include 6.5x more tokens at Firmus, 4.2x effective GPU capacity at CoreWeave, and a published 10x token improvement on Oracle Cloud. AMG combines the WEKA storage platform with integrations into inference stacks such as Dynamo, TensorRT-LLM, vLLM, and LM Cache, using GPU-server NICs for up to 3.2 terabits per second of KV cache traffic. Ben David also describes an exabyte-scale customer cluster sustaining over 13 terabytes per second and 100 million IOPS, driven by an autonomous vehicle company running training, inference, and data analytics, and discusses the layered neocloud market where providers often cannot predict tenant workloads.