Dell Pro Max GB10 enables local AI clusters; dual Blackwell systems cluster for 256GB shared VRAM

Dell's Pro Max with NVIDIA GB10 enables cost-effective local AI clustering through NVIDIA's ConnectX 7 200Gbps NICs and Remote Direct Memory Access over Converged Ethernet (RoCE). A pair of Pro Max GB10 systems, each with 128GB of LPDDR5X unified memory and 6,144 CUDA cores on Blackwell GPU, can be networked together to create a 256GB local AI sandbox, tested by Tom's Hardware.

At roughly $6,332 per system for the tested 4TB SSD configuration, a dual-system cluster costs substantially less than building an equivalent GPU server with four discrete 48–72GB cards plus the required CPU, motherboard, and DDR5 platform—which alone demands a high-end Threadripper Pro or EPYC socket and exceeds standard 15-amp home power circuits. The Pro Max trades some absolute performance for accessibility and practicality: no single powerful host is required, and power consumption remains manageable in shared office or lab spaces.

The GB10's architecture preserves the DGX Spark reference design's clustering-first approach. Each Pro Max includes a 280W USB-C adapter, PCIe Gen 4 NVMe SSD (Gen 5 availability constrained by the chip shortage), and ConnectX 7 configured for Ethernet + RoCE only—not Infiniband. This RoCE-over-Ethernet backbone enables distributed inference workloads across both systems as a unified compute pool.

For researchers, startups, and engineers needing local inference on larger open models without data-center cloud spend, the Pro Max GB10 pair removes a major friction point: fitting multi-hundred-billion-parameter models into local silicon at sub-$13K total cost with professional clustering networking. The simplicity of USB-C power and IP networking beats bespoke GPU server builds, though absolute throughput remains below rack-scale solutions.