GIGABYTE NVIDIA Vera Rubin NVL72 GIGAPOD AI DLC Rack

The GIGABYTE NVIDIA Vera Rubin NVL72 GIGAPOD AI DLC Rack represents the absolute pinnacle of rack-scale AI computing, bringing NVIDIA's next-generation Vera Rubin architecture into a fully integrated, liquid-cooled enterprise deployment. Engineered for massive foundational model training, high-throughput inference, and deep reasoning agentic workflows, the system packages 72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs into a unified 1U-tray-based footprint. Often classified under enterprise and custom-built hardware directories, this platform is built specifically for data centers, frontier model labs, and sovereign AI initiatives requiring extreme single-rack compute density.

At its core, the GIGAPOD AI DLC rack functions not as an isolated cluster of independent servers, but as a single coherent supercomputer. By linking 72 GPUs across sixth-generation NVLink with 260 TB/s of scale-up bidirectional bandwidth, the system eliminates traditional node-to-node PCIe and networking bottlenecks during tensor-parallel execution. GIGABYTE showcased this complete, fully integrated production unit at the OCP Global Summit on October 15, 2026, delivering an turnkey answer to the physical and thermal demands of gigawatt-scale AI infrastructure.

For teams deploying the GIGABYTE NVIDIA Vera Rubin NVL72 GIGAPOD AI DLC Rack for AI workloads, the platform sets a new efficiency benchmark. NVIDIA rates the architecture to deliver AI inference at one-tenth the cost per million tokens and up to ten times more tokens per megawatt compared to prior Blackwell GB200 NVL72 installations, specifically when handling deep reasoning models and long-context agentic operations.

The compute specifications of the NVL72 platform redefine scale-up capacity in a single 1,200 x 600 x 2,300 mm footprint. Powered by 18 GIGABYTE XN16-CB0-L01 1U liquid-cooled compute trays, each tray houses two NVIDIA Vera CPUs (each packing 88 custom Olympus cores) and four NVIDIA Rubin GPUs.

• GPU Compute: 72 x NVIDIA Rubin GPUs delivering 288,000 TFLOPS (288 PFLOPS) of dense FP16/BF16 compute, 3,600 PFLOPS of NVFP4 inference, and 18,000 TOPS (18 POPS) of INT8 throughput.
• CPU Compute: 36 x NVIDIA Vera CPUs with 3,168 total Olympus ARM cores backed by 54 TB of high-speed LPDDR5X memory.
• GPU Memory: 20,700 GB (20.7 TB) of HBM4 memory distributed across the 72 GPUs at 288 GB per GPU.
• Memory Bandwidth: 1,400,000 GB/s (1,400 TB/s) aggregate HBM4 bandwidth across the full rack.
• Interconnect: Sixth-generation NVIDIA NVLink delivering 260 TB/s of aggregate scale-up bandwidth via 9 NVLink switch trays and 4 cable cartridges.
• Scale-Out Networking: 8 x 800 Gb/s OSFP ports via NVIDIA ConnectX-9 SuperNICs and 2 x 400 Gb/s QSFP ports via NVIDIA BlueField-4 DPUs per compute tray, integrating directly with Quantum-X800 InfiniBand or Spectrum-X Ethernet switches.
• Power and Cooling: Direct Liquid Cooling (DLC) with an internal distribution manifold designed for in-row CDUs, powered by four 3RU 110 kW power shelves backed by a 54V DC, 5,000A copper busbar.

Memory capacity and memory bandwidth dictate raw token generation speeds in autoregressive LLMs. The 1.4 PB/s of aggregate memory bandwidth ensures that memory-bound decoding phases execute without compute starvation, even when processing massive key-value (KV) caches across hundreds of concurrent user sessions.

With a 20,700 GB GPU for AI memory pool linked over NVLink, the question is not which single model fits, but rather how many frontier-scale models the rack can host concurrently in unquantized precision.

• DeepSeek-R1 and DeepSeek-V3 (671B MoE): The entire 671-billion parameter model fits directly into HBM4 memory in native 16-bit precision (requiring approximately 1.3 TB to 1.5 TB of VRAM including runtime context). The NVL72 can run multiple instances of DeepSeek-R1 at FP16 simultaneously, or run a single instance across all 72 GPUs using tensor and pipeline parallelism to hit ultra-low time-to-first-token (TTFT) metrics.
• Llama 3.1 405B: Meta's 405B dense model consumes roughly 810 GB in FP16 and 405 GB in FP8. On this system, a full FP16 instance requires less than 4 percent of total rack VRAM, allowing enterprise teams to run production clusters of 405B models alongside multi-million token KV caches without offloading.
• Frontier Multi-Trillion Parameter Systems: The 20.7 TB memory capacity provides native headroom for emerging 1T to 2T parameter architectures in 8-bit and 4-bit precision formats on a single rack, avoiding the high cross-rack latency typical of multi-node InfiniBand clusters.

For teams searching for hardware for running 70B parameter models, the NVL72 operates as a high-density token factory. A model like Llama 3.3 70B or Qwen 2.5 72B occupies roughly 140 GB in FP16 or 70 GB in FP8. A single NVL72 rack can host dozens of isolated 70B model replicas concurrently, delivering tens of thousands of aggregate tokens per second across parallel enterprise workloads.

• NVFP4 (Microscopic Precision): Rubin introduces optimized 4-bit floating point inference hardware routines (3,600 PFLOPS rack throughput). This is the optimal configuration for real-time agentic reasoning loops, providing maximum throughput per watt while maintaining accuracy on frontier reasoning architectures.
• FP8/FP6: The ideal sweet spot for production training and high-fidelity enterprise serving, balancing memory footprint with zero degradation in mathematical stability.
• Native FP16/BF16: Reserved for foundation model pre-training, complex synthetic data generation, and mathematical validation benchmarks.

The GIGABYTE NVL72 is strictly a data-center-class solution requiring industrial three-phase power and facility chilled water loops conforming to ASHRAE liquid cooling standards. It is tailored for distinct enterprise workflows:

• Agentic AI Orchestration: Real-time multi-agent systems rely on deep reasoning loops, high context retention, and rapid execution. The NVL72 delivers the low inference latency and massive memory headroom necessary to support agentic swarms running parallel reasoning passes.
• Foundational Model Pre-Training and Post-Training: With 288 PFLOPS of FP16 compute and 260 TB/s NVLink bandwidth, research labs can train and fine-tune large foundation models with one-fourth the physical GPU footprint compared to previous hardware architectures.
• Large-Scale Inference Infrastructure: Cloud service providers and AI API vendors handling millions of daily requests benefit from the high token density per megawatt, lowering unit costs per generated token.

This hardware is not meant for individual desktop users or edge deployments. It is engineered for enterprise IT teams, high-performance computing (HPC) centers, and AI infrastructure architects deploying centralized compute hubs.

The prior-generation GB200 NVL72 established the liquid-cooled rack standard using Blackwell B200 GPUs (192 GB HBM3e) and Grace CPUs. The Vera Rubin NVL72 upgrades every core subsystem:

• Memory Density: Rubin increases VRAM to 288 GB HBM4 per GPU (20.7 TB per rack) compared to 192 GB HBM3e (13.8 TB per rack) on GB200.
• Interconnect: Sixth-generation NVLink boosts aggregate bandwidth to 260 TB/s, doubling communication speeds for tensor-parallel communication.
• Efficiency: For deep-reasoning agentic workloads, the Rubin architecture delivers up to a 10x reduction in cost per million tokens and significantly higher token throughput per megawatt.

While AMD offers competitive individual accelerators with high HBM capacities (such as 192 GB to 256 GB per GPU), AMD configurations typically rely on 8-GPU nodes linked together over standard Ethernet or InfiniBand fabrics. The GIGABYTE NVL72 differentiates itself via its unified 72-GPU NVLink switch fabric. Distributing a single 1-trillion parameter model across 72 GPUs over 260 TB/s NVLink results in dramatically lower all-reduce latencies than splitting the same workload across discrete PCIe or NIC-connected servers. For large-scale distributed inference and massive context windows, the NVL72 remains the superior architectural standard.