NVL72 Portfolio
Two Rack-Scale Platforms. Different Strengths.
Both platforms scale 72 GPUs inside a single NVLink domain and scale out over NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet. The difference is the generation of compute, memory, networking, and the workloads each platform is designed to optimize.
Blackwell Ultra • Available Now
GB300 NVL72
Built for AI reasoning scaling
A production-ready rack-scale platform combining 72 Blackwell Ultra GPUs and 36 Grace CPUs with fifth-generation NVLink and ConnectX-8 networking.
72
Blackwell Ultra GPUs
20 TB
HBM3E GPU memory
130 TB/s
NVLink bandwidth
Explore GB300 NVL72
Vera Rubin • NEXT GENERATION AI
Vera Rubin NVL72
Built for the agentic AI factory
NVIDIA’s next-generation rack-scale platform combines 72 Rubin GPUs and 36 Vera CPUs with NVLink 6, ConnectX-9 SuperNICs, BlueField-4 DPUs, and HBM4 memory.
72
Rubin GPUs
20.7 TB
HBM4 GPU memory
216 TB/s
NVLink bandwidth
Explore Vera Rubin NVL72
NVIDIA Rack-Scale AI Systems
Choose the Right NVL72 for Your AI Factory
From Blackwell Ultra to Vera Rubin, ADS integrates NVIDIA’s 72-GPU rack-scale platforms with the networking, storage, liquid cooling, facility planning, validation, and lifecycle support required for production AI.
Two generations. Two different deployment profiles. One integration partner focused on matching the architecture to your workload, timeline, and data center.
Request a Consultation
At a Glance
GB300 vs. Vera Rubin NVL72
A high-level comparison of the two NVIDIA rack-scale generations.
Platform Attribute
GB300 NVL72
Vera Rubin NVL72
Primary Positioning
AI reasoning, scaling inference, training, and HPC
Agentic AI, advanced reasoning, MoE training, large-context inference, AI + HPC
GPU Configuration
72 NVIDIA Blackwell Ultra GPUs
72 NVIDIA Rubin GPUs
CPU Configuration
36 NVIDIA Grace CPUs
36 NVIDIA Vera CPUs
GPU Memory
20 TB HBM3E
20.7 TB HBM4
GPU Memory Bandwidth
Up to 576 TB/s aggregate
1,400 TB/s aggregate per NVIDIA reference specifications
Scale-Up Interconnect
Fifth-generation NVLink
Sixth-generation NVLink
NVLink Bandwidth
130 TB/s
216 TB/s
Scale-Out Networking
ConnectX-8; Quantum-X800 InfiniBand or Spectrum-X Ethernet
ConnectX-9 + BlueField-4; Quantum-X800 InfiniBand or Spectrum-X Ethernet
CPU Memory
17 TB LPDDR5X
Up to 54 TB LPDDR5X
NVIDIA Availability Positioning
Available now
Started production
Which Platform Fits?
Match the Generation to the Deployment
GB300 NVL72 is a fit when…
✓
You need production Blackwell Ultra infrastructure now.
✓
Your priority is AI reasoning, test-time scaling, inference, or large-model training.
✓
Your facility and operating model are already aligned to the current GB300 ecosystem.
✓
You want a mature, deployable NVL72 architecture with current NVIDIA Mission Control support.
✓
You need to scale from one NVL72 rack into a larger AI factory using 800G networking.
ADS can help validate workload fit rather than forcing a generation decision from specifications alone.
Vera Rubin NVL72 is a fit when…
✓
Your roadmap is centered on next-generation agentic AI and higher token volumes.
✓
Large-context workloads and increasingly large mixture-of-experts models drive the architecture.
✓
HBM4 bandwidth and NVLink 6 scale-up performance materially affect your workload economics.
✓
You need NVIDIA's newest rack-scale confidential-computing architecture.
✓
You are designing the data center around the next generation rather than fitting AI into an existing footprint.
Vera Rubin raises the importance of facility power, liquid cooling, networking, and data architecture planning.
The ADS Difference
The Rack Is Only the Beginning
The value of an NVL72 system depends on whether the rest of the infrastructure can keep it productive. ADS architects the system around the rack—not the other way around.
Workload-First Design
Size the platform around reasoning, training, inference, agentic workflows, scientific computing, and real data movement requirements.
High-Speed Fabric
Engineer scale-out networking around Quantum-X800 InfiniBand or Spectrum-X Ethernet and the NIC/DPU configuration of the chosen NVL72 generation.
ADS Storage Portfolio
Match IBM Storage Scale, BeeGFS, VDURA, or custom Ceph to checkpointing, model data, context, object, file, and capacity requirements.
Facility Readiness
Plan power, CDU capacity, liquid loops, water requirements, rack placement, heat rejection, and network uplinks before delivery.
Integration & Validation
Coordinate compute, fabric, storage, cooling, software, and management so the environment is validated as a system before production.
One Accountable Partner
ADS provides a single integration partner focused on the end-to-end outcome instead of leaving customers to coordinate multiple infrastructure vendors.
Don’t just choose an NVL72. Build the AI factory around it.
ADS helps turn rack-scale NVIDIA compute into a balanced, deployable, and supportable production platform.
Integrated Architecture
Compute → Fabric → Data
Whether you choose GB300 or Vera Rubin, the surrounding architecture determines how effectively the GPUs receive data, communicate across racks, checkpoint models, and serve production workloads.
NVL72 Compute
GB300 Blackwell Ultra or Vera Rubin rack-scale compute.
→
High-Speed AI Fabric
NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet.
→
ADS Storage
IBM Storage Scale • BeeGFS • VDURA • Custom Ceph
Not Sure Which NVL72 Fits Your Roadmap?
Bring us your workload, timing, data profile, scale target, networking requirements, and facility constraints. ADS can help evaluate the complete system—not just the GPU generation.
Request a Consultation
Applied Data Systems
©2026 Applied Data Systems