DDN today announced a deep strategic partnership with NVIDIA to power the next-generation AI factory architecture showcased at CES 2026, integrating DDN’s AI data platform with the NVIDIA Rubin compute fabric and BlueField-4 DPU. The joint solution delivers up to 99% sustained GPU utilization across multi-rack installations, slashes time-to-first-token (TTFT) by 20–40% on million-token context inference workloads, and reduces host CPU overhead by up to 30% through intelligent offload of metadata processing and storage services onto BlueField-4. Enterprises and hyperscalers running long-context and distributed inference at scale can now deploy models faster, maintain predictable performance under peak concurrency, and lower total cost of ownership by eliminating inefficient data movement between storage, network and accelerators.

At CES 2026, NVIDIA introduced its Vera Rubin AI platform—a rack-scale architecture that co-designs CPUs, GPUs, DPUs and high-speed NVLink fabrics to unlock exascale AI throughput in a unified system. Paired with Spectrum-X Ethernet, Vera Rubin supports line-rate data delivery into dense GPU arrays and implements a distributed KV cache tier optimized for long-memory inference. NVIDIA also debuted Alpamayo, an open-source foundation model designed for autonomous vehicles, featuring a 2 billion–parameter vision-language encoder and optimized training recipes that reduce model convergence time by 25%. Alpamayo’s modular design allows OEMs to finetune on real-world telematics datasets in under 48 hours, accelerating deployment of perception and decision-making stacks in production fleets.

In a CNBC interview at the show floor, NVIDIA founder and CEO Jensen Huang framed modern supercomputers as “AI factories” and spotlighted the Vera Rubin AI chip, which delivers a four-fold performance improvement over its predecessor, Blackwell. Huang explained that this leap enables data centers to train equivalent models in one-quarter the time or to consolidate hardware footprints by replacing four Blackwell nodes with a single Vera Rubin module. This efficiency gain translates into substantial savings in power consumption—approximately 60% lower at peak throughput—and frees capital for reinvestment in network fabrics and data-management layers critical for enterprise-grade AI deployments.