What it means to solve the GPU data access problem
Solving the GPU data access problem means giving enterprises fast, secure GPU cloud infrastructure that can reach data wherever it lives, without endless copying, replication, or manual staging steps that stall AI projects and leave GPU resources idle. For years, the gap between powerful accelerators and slow-moving enterprise data has held back AI adoption at scale. A recent analysis cited by Qumulo shows average enterprise GPU utilization at about 5%, which means expensive GPU clusters sit idle most of the time while teams move datasets into place. New enterprise data fabric designs aim to flip this model: instead of chasing GPUs and duplicating storage in every cloud, they present a single, consistent data view to many GPU pools, with multi-region GPU access and AI data management handled behind the scenes.
Qumulo’s AI data fabric: connecting GPUs to live enterprise data
Qumulo’s Cloud AI Accelerator is built around an enterprise data fabric that links distributed data directly to GPU resources across clouds, regions, and hybrid environments. Rather than replicating petabytes into every GPU cluster, Qumulo Cloud Data Fabric, Cloud Native Qumulo, and NeuralCache present a real-time, consistent view of files to any connected GPU farm. This architecture eliminates the heavy load phase into GPU-attached flash that often takes weeks before training or inference can begin. Qumulo describes this as creating “GPU liquidity”: workloads run wherever GPU capacity is available, instead of where data is trapped in storage islands. Enterprises can connect on-premises or cloud-native Qumulo systems to services such as Microsoft AI Foundry, AWS Bedrock, and Google Vertex AI without copying data, reducing idle compute and simplifying AI data management across multi-cloud environments.
SoftBank’s Infrinia GPU cloud and sovereign AI compute
SoftBank’s AI Data Center GPU Cloud, powered by Infrinia AI Cloud OS and based on NVIDIA GB200 NVL72 systems, tackles a different but related bottleneck: how to offer high-density, centrally managed GPU cloud infrastructure with strict control over where data and compute reside. Infrinia AI Cloud OS provides Kubernetes as a Service for multi-tenant clusters and Inference as a Service for API-based large language model workloads, so enterprises do not have to assemble a custom stack. According to SoftBank, the platform lets one pooled GPU resource handle both intensive training and latency-sensitive inference, managed from a single control layer. By keeping GPU capacity and operational software under one roof, this design supports sovereign AI workloads and keeps latency low for users who need AI compute within a defined jurisdiction while maintaining clear AI data management boundaries.
Why data fabrics and GPU clouds matter for multi-region AI
The core challenge for large-scale AI is no longer only how much GPU capacity an enterprise can secure, but how quickly GPU cloud infrastructure can see and use the right data. Qumulo’s data fabric attacks the data-gravity side of the equation, so GPUs spread across regions and clouds can work on live datasets without replica sprawl or staging delays. SoftBank’s Infrinia AI Cloud focuses on turning big, specialized GPU clusters into a shared, policy-controlled pool that can power many AI workloads through Kubernetes orchestration and inference APIs. Together, these approaches point toward multi-region GPU access that does not depend on moving data to every cluster. Enterprises gain the ability to schedule training and inference where GPUs are free, while keeping a single source of truth for data and a clearer operational model for AI data management.