Qumulo and Nvidia support - Blocks & Files

In the realm of AI and machine learning, data processing efficiency is a critical determinant of model training speed. Recently, NVIDIA introduced the STX Key-Value (KV) caching scheme to optimize dat... Read original →

Qumulo’s refusal to adopt NVIDIA’s STX isn’t technical obstinacy—it’s a strategic bet on data gravity over GPU proximity. Technically, this decouples storage from proprietary GPU caching, pushing CNQ and NeuralCache toward becoming cross-cloud data plane standards and pressuring NVIDIA to open BlueField-4/Spectrum-X interfaces. From a compliance angle, vendor lock-in in KV caching heightens supply chain fragility amid U.S. export controls on AI chips to China; true multi-cloud data liquidity is now a risk-mitigation imperative. Competitors like VAST Data will likely counter with their own zero-staging architectures, eroding NVIDIA’s control over the storage-compute boundary. Over the next 18 months, as >70% of AI workloads run in public clouds, training velocity will hinge not on FLOPS but on deterministic, low-latency data delivery across regions—making Qumulo’s focus on eliminating staging delays the decisive bottleneck to solve.