DDN and Nvidia team up to let GPUs grab their own data, targeting AI’s biggest bottleneck

Here’s a dirty secret about AI infrastructure: some of the most expensive hardware on the planet spends a surprising amount of time doing absolutely nothing. GPUs, the workhorses of modern AI training and inference, frequently sit idle while waiting for data to arrive from storage systems. DDN and Nvidia are now collaborating to fix that problem by letting GPUs initiate data access directly, cutting out the middleman.

The partnership centers on DDN’s Infinia data intelligence platform, which integrates with Nvidia’s Storage-Next initiative to enable GPU-initiated data operations. In English: instead of GPUs politely waiting for CPUs to fetch and deliver data, the GPUs can now reach out and grab what they need themselves.

What the technology actually does

The result is reduced latency, minimized CPU overhead, and significantly better GPU utilization. For organizations running large-scale AI factories, where thousands of GPUs operate in concert, even marginal improvements in utilization translate into meaningful cost savings.

DDN CTO Sven Oehme framed the collaboration in economic terms, pointing to improvements in “the economics of AI at scale.” Jason Hardy, Nvidia’s VP of Storage Technology, emphasized the seamless integration between computing and data infrastructure as a key advantage.

The technology fits into Nvidia’s broader 2026 platform strategy, which includes the Nvidia Rubin platform and BlueField-4 DPUs.

A partnership a decade in the making

DDN and Nvidia have been working together since 2016, when DDN became the first storage vendor to certify its platform with Nvidia’s GPUDirect Storage technology.

GPUDirect Storage allowed data to move directly between storage and GPU memory without bouncing through system memory first. The new collaboration with Infinia takes that concept further by giving GPUs the ability to actively initiate those data transfers rather than passively receiving them.

DDN plans to demonstrate these capabilities at the Flash Memory Summit, where the company is expected to share performance benchmarks and details about operational availability.

Why this matters for the AI infrastructure market

As models grow larger and training datasets expand, the gap between compute capability and data delivery widens. Large AI training runs can involve petabytes of data spread across distributed storage systems. Every microsecond of latency in data access compounds across thousands of GPUs and millions of training iterations.

Rather than simply making storage faster, DDN and Nvidia are rethinking how data flows through the system entirely. Giving GPUs direct agency over data operations represents a fundamental shift in how AI infrastructure operates.