Dell introduces PowerEdge XE8812, a new system architecture optimized for NVIDIA Vera Rubin NVL4, enabling up to 144 GPUs per rack.

Dell Technologies introduced the PowerEdge XE8812 server at ISC 2026, built around NVIDIA's Vera Rubin NVL4 architecture. The XE8812 is offered in the OCP standards-based Dell PowerRack 9100 and supports up to 144 GPUs per rack, with rack-level power delivery exceeding 300 kW. The platform features a fanless, direct liquid-cooled design targeting large-scale inference, scientific simulation, and foundation-model workloads.

Compared to the prior generation, the XE8812 increases host CPU core counts from 144 to 176 and provides approximately 50% more memory per socket and GPU memory. The direct liquid cooling design for both CPUs and GPUs enables higher density and thermal efficiency for rack-scale deployments. The Dell PowerRack 9100 serves as an OCP-compatible rack architecture to house the dense XE8812 nodes and associated power and telemetry systems.

Industry practitioners provisioning rack-scale AI and HPC infrastructure have prioritized tight integration of high-bandwidth GPUs, larger host memory footprints, and liquid cooling to keep models and datasets resident in memory and reduce I/O-induced latency. The Vera Rubin platform emphasizes dense GPU counts per rack and multi-exaflop aggregate performance, aligning with the XE8812's design trade-offs toward compute density and thermal management. These trade-offs typically increase demands on cluster power distribution, chilled-water infrastructure, and network fabric to sustain peak utilization.

NVIDIA framed Vera Rubin at ISC 2026 as delivering more than multi-exaflops of AI performance and architectures enabling up to 144 GPUs per rack. Dell and Super Micro were named as system builders for the platform. Dell positions the XE8812 as part of the Dell AI Factory with NVIDIA initiative and ties the product to use cases in genomics, engineering, and sovereign AI efforts. Industry reporting indicates investor interest in vendors participating in Vera Rubin NVL4 deployments.

Practitioners assessing this platform should track vendor availability windows and qualification details for GPU SKUs and interconnect configurations, since aggregated rack GPU counts and power envelopes are sensitive to chosen GPU memory sizes and network topologies. Monitor partner ecosystem updates for software stack integration (CUDA-X, MPI/collectives) and coolant infrastructure requirements, as dense liquid-cooled racks typically require integration work at the data-center level.