Expanding AI Storage Access with NVIDIA cuObject and the NVIDIA SCADA Server SDK

NVIDIA says cuObject is generally available with standardized APIs and an RDMA wire protocol so accelerated object access needn’t bounce through the server CPU. The SCADA Server SDK lets storage vendors answer GPU-initiated requests; IBM Storage has shown a prototype with Storage Scale.

The xio-sig consortium is expanding to include cuObject alongside cuFile, with Google Cloud evaluating participation and Microsoft planning to join the board. Storage-Next coordinates 40-plus vendors and customers around open GPU-driven fine-grained storage access. cuObject Server 2.0.0 is available to evaluate.

We’re glad storage is finally being treated as part of the AI critical path.

GPUs starving on CPU-mediated object fetches waste expensive silicon. An open-ish wire protocol with cloud vendors at the table beats twenty proprietary RDMA dialects. Useful AI factories need data movement as engineered as compute.

I’m watching whether Google and Microsoft actually ship interoperable endpoints or just take board seats. Standards die in press releases.

We’ll track IBM’s prototype and early cuObject adopters. Zero-copy object I/O is how inference context and training pipelines stay sane at scale.

NVIDIA Technical Blog, September 30, 2026, cuObject GA and SCADA Server SDK.