Rack-scale AI infrastructure for the era of reasoning — 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs in a single NVLink-connected, liquid-cooled rack.
Manufacturer-defined features from the published datasheet.
NVIDIA GB300 NVL72 connects 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs in a rack-scale, liquid-cooled design. All 72 GPUs are interconnected via fifth-generation NVLink and present a coherent ~21 TB HBM3e memory pool.
Per NVIDIA's published figures, NVIDIA GB300 NVL72 delivers approximately 1.8× the inference throughput and 1.5× the training performance of the prior-generation NVIDIA GB200 NVL72.
Per NVIDIA's published figures, up to 1.4 ExaFLOPS of dense FP4 inference performance per rack. The architectural primitive behind NVIDIA's AI Factory reference designs.
Fifth-generation NVLink with NVLink Switch delivers 130 TB/s aggregate intra-rack bandwidth — all 72 GPUs interconnected all-to-all, coherent memory addressing across the rack.
Cluster-level orchestration platform from NVIDIA — health monitoring, telemetry aggregation, validated recipe deployment, job-scheduler integration. Standard operating model for NVIDIA DGX SuperPOD with GB300.
Multi-rack scale-out via NVIDIA ConnectX-8 SuperNICs at 800 Gb/s, paired with NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X800 Ethernet rack fabrics.
The NVIDIA HGX baseboard is identical on both paths. The turnkey DGX is fastest to deploy; OEM HGX platforms (Dell, Giga Computing, Supermicro) give wider configuration choice.
NVIDIA-built and NVIDIA-supported packaging of the NVIDIA GB300 NVL72 platform. Ships pre-configured with NVIDIA DGX OS, NVIDIA AI Enterprise, NVIDIA Mission Control, and a multi-year NVIDIA Enterprise Support contract.
The same NVIDIA GB300 NVL72 rack platform from NVIDIA's OEM partners — customer-operated, no DGX software bundle. Same NVIDIA Grace CPUs, NVIDIA Blackwell Ultra GPUs, and NVLink Switch fabric. OEM warranty and support model.
EMARQUE supplies both paths in Malaysia. Final configuration, lead time, and warranty terms are confirmed in writing at quotation.
The four sub-systems that determine real-workload behaviour. We tune each before delivery.
Get a GB300 NVL72 configuration and lead time from your Malaysian NVIDIA systems specialist.
Workload categories documented in the manufacturer's reference materials. Sizing is confirmed with your technical team during scoping.
NVIDIA positions GB300 NVL72 for the AI reasoning era. The coherent ~21 TB HBM3e memory pool supports extended-context inference and multi-step agentic workloads at production scale.
Single-rack tensor parallelism across 72 NVLink-connected GPUs supports trillion-parameter foundation model training without cross-rack communication penalties for in-rack workloads.
NVIDIA GB300 NVL72 is the building block for NVIDIA DGX SuperPOD with GB300 systems. Multi-rack scale-out via NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X800 Ethernet.
NVIDIA references the NVL72 rack as the standard planning unit for sovereign-compute and national-lab AI infrastructure projects. EMARQUE handles in-country project delivery, commissioning, and Tier-1 support for multi-rack deployments.
As supplied by NVIDIA. EMARQUE handles in-country delivery, commissioning, and Tier-1 support handoff.
Per NVIDIA: a rack-scale AI compute platform that connects 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs in a single liquid-cooled rack via fifth-generation NVLink and NVLink Switch. All 72 GPUs operate as a coherent NVLink Switch domain with approximately 21 TB of HBM3e memory exposed as one pool. It is the building block for NVIDIA DGX SuperPOD with GB300 systems.
Per NVIDIA: approximately 1.4 ExaFLOPS of dense FP4 inference performance per rack, approximately 1.8× the inference throughput and 1.5× the training performance of the prior-generation NVIDIA GB200 NVL72, and 130 TB/s aggregate NVLink bandwidth within the rack.
NVIDIA DGX B300 is a single-node 8-GPU air-cooled system in a 10U form factor. NVIDIA GB300 NVL72 is a rack-scale 72-GPU liquid-cooled platform. The NVLink Switch fabric in NVIDIA GB300 NVL72 interconnects all 72 GPUs as a single coherent accelerator from a tensor-parallelism perspective; NVIDIA DGX B300 nodes connect to each other via NVIDIA Quantum-X800 InfiniBand at the cluster layer.
Liquid-cooled rack — CDU (Coolant Distribution Unit) integration required. Approximately 120 kW per rack typical load. Three-phase high-amperage power circuits with appropriate PDU planning. Inlet liquid temperature 30–35 °C per NVIDIA's reference design (warm-water cooling capable). Raised-floor or in-row cooling configuration. EMARQUE conducts a full site readiness assessment as part of project scoping.
Both use the same NVIDIA GB300 NVL72 rack platform — same NVIDIA Grace CPUs, same NVIDIA Blackwell Ultra GPUs, same NVLink Switch fabric, same liquid-cooled rack. NVIDIA DGX GB300 is the NVIDIA-branded packaging delivered with NVIDIA DGX OS, NVIDIA Mission Control orchestration, and a multi-year NVIDIA Enterprise Support contract. OEM versions (Giga Computing, Supermicro, Dell, ASUS) ship the same hardware platform with the OEM's warranty and support model; NVIDIA AI Enterprise software is available separately. Both options are presented on this page.
Lead time follows NVIDIA's allocation schedule for NVIDIA GB300 NVL72 platforms. Multi-rack NVIDIA DGX SuperPOD projects are scoped individually — typical timeline spans allocation reservation, freight, in-country commissioning, fabric build-out, and acceptance testing. EMARQUE confirms projected delivery window at order acknowledgement following NVIDIA allocation.
NVIDIA's announced next-generation rack-scale platform is NVIDIA Vera Rubin NVL72, targeted for the late-2026 / 2027 window — same NVL72 rack architecture with the NVIDIA Vera CPU and NVIDIA Rubin GPU generation, plus HBM4 memory. EMARQUE accepts allocation conversations for NVIDIA Vera Rubin NVL72 for customers planning multi-year refresh cycles.
Manufacturer specifications, factory lead times, and warranty terms apply. EMARQUE responds within one business day with a formal quotation and projected delivery window.
Tell us about your workload. We reply within one business day with a quote sized to fit.