From Data Gravity to GPU Liquidity
GPU cloud platforms are shared computing environments that provide accelerated processors, storage, and networking as a service so organizations can run demanding AI workloads without owning physical infrastructure, while minimizing data movement delays that slow model training and inference. The main bottleneck they now target is data gravity: models wait while data is copied, staged, and replicated to wherever GPUs reside. This delay translates into idle accelerators, low utilization, and slow time-to-results. A new wave of enterprise GPU infrastructure focuses on the opposite pattern: keep data where it is, and route GPU access to it. That means connecting file and object stores directly to GPU clusters, coordinating access across multiple clouds, and automating placement decisions so AI workloads can start as soon as GPUs are free, instead of after data pipeline catch-up.
Qumulo’s AI Data Fabric and the End of Staging Delays
Qumulo’s Cloud AI Accelerator is positioned as an AI data fabric that links enterprise datasets to GPU cloud platforms across regions and providers without replication or staging. According to Qumulo’s analysis, the average enterprise GPU utilization hovers around 5%, largely because data must be staged and replicated before workloads start. Qumulo addresses this by presenting distributed data in real time to GPU farms in any cloud, turning “GPU hunting” from a logistical task into a scheduling decision. Its stack combines Cloud Native Qumulo, Cloud Data Fabric, and NeuralCache to create what the company calls enterprise GPU liquidity: workloads run wherever GPU capacity exists, rather than where data is trapped. By enabling direct GPU access to enterprise data, the platform reduces GPU access latency, shrinks idle time, and helps teams use existing cloud GPU deployments more efficiently.
SoftBank’s Infrinia AI Cloud OS and Sovereign GPU Cloud
SoftBank’s AI Data Center GPU Cloud adds another piece to the puzzle by pairing high-density NVIDIA GB200 NVL72 systems with Infrinia AI Cloud OS. The software layer offers Kubernetes as a Service and Inference as a Service so users can run training and latency-sensitive inference on a single GPU pool. This design supports multi-tenant workloads with centralized, automated management of GPU resources, helping enterprises avoid fragmented stacks for different AI tasks. SoftBank also emphasizes sovereignty, promising AI compute that stays within defined jurisdictions and closer to the network edge. This aligns with regulatory pressure on cross-border data flows and the desire to keep sensitive training data local. By integrating sovereign compute, GPU orchestration, and inference APIs, the platform aims to reduce end-to-end GPU access latency from data center to edge while keeping control in the customer’s hands.
Virtuozzo’s Hyperconverged Infrastructure: AI Built With AI
Virtuozzo’s Infrastructure System introduces a hyperconverged approach to enterprise GPU infrastructure: compute, storage, networking, and a next-generation operating system in a single architecture. The company describes its vision as AI infrastructure built with AI, by AI, and for AI, focusing on efficiency and cost. Its V/OS offers a tuned Linux foundation that supports both virtual machines and system containers, aiming for near bare-metal performance and dense GPU cloud platform deployments. V/Orchestration, V/Management, V/Automation, and V/Protection combine to remove fragmentation: unified control of Kubernetes environments, built-in billing and provisioning, and integrated backup and security. Virtuozzo claims this can deliver 60–80% lower total cost of ownership compared with typical hosting setups. For AI teams, the payoff is consistent performance and fewer silos between compute and storage, which in turn reduces GPU access latency and makes it easier to place workloads where GPUs and data align.
Why Direct Data-to-GPU Paths Matter for AI Performance
Across these platforms, a common pattern is emerging: direct data-to-GPU paths instead of slow, repeated data movement. Qumulo’s AI data fabric connects file systems and object stores to GPUs wherever they sit, SoftBank’s Infrinia AI Cloud OS orchestrates training and inference across a unified GPU pool, and Virtuozzo’s hyperconverged stack keeps compute and storage under one roof. Together, these trends attack the root cause of low GPU utilization and delay: the need to copy and restage large datasets every time workloads move. By shortening the gap between data and compute, enterprises can treat GPUs as a fluid resource across clouds and regions, not a fixed asset tied to one data silo. The result is faster AI model training and inference, better use of expensive accelerators, and cloud GPU deployments that respond in minutes instead of days when demand shifts.