NVIDIA Unveils the Vera Rubin NVL72: Full Specs and Platform Breakdown

NVIDIA announced the Vera Rubin platform at GTC 2026. Vera Rubin is a seven-chip, five-rack AI supercomputer architecture designed for agentic AI workloads. The platform is in full production as of Q1 2026 with partner availability in H2 2026.

TLDR

• Vera Rubin NVL72 packs 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled rack delivering 3.6 EFLOPS of NVFP4 inference and 2.5 EFLOPS of training compute.
• Seven co-designed chips span compute, networking, storage, and security—including the Groq 3 LPU from NVIDIA's $20B acqui-hire of Groq.
• NVIDIA claims 5x inference performance and 10x lower cost per token versus Blackwell at the rack level.
• A full Vera Rubin POD scales to 40 racks, 1,152 GPUs, and 60 exaflops.
• First deployments expected from AWS, Google Cloud, Microsoft, OCI, and CoreWeave in H2 2026.

Platform Overview

The Vera Rubin platform comprises seven chips across five rack-scale systems. Each chip serves a dedicated function across compute, networking, and data management.

The Seven Chips

Below is a detailed breakdown of the specifications for each of the seven co-designed chips that form the Vera Rubin platform.

Rubin GPU

• Process: TSMC 3nm, dual-die design
• Transistors: 336 billion
• Memory: 288 GB HBM4, 22 TB/s bandwidth
• Inference Performance: 50 PFLOPS NVFP4 (5x Blackwell)
• Training Performance: 35 PFLOPS NVFP4 (3.5x Blackwell)
• Transformer Engine: 3rd generation with hardware-accelerated adaptive compression
• Power Profiles: Max Q (~1.8 kW TGP) and Max P (~2.3 kW TGP)

Vera CPU

• Cores: 88 custom Arm-based Olympus cores (Armv9.2), 176 threads via Spatial Multithreading
• Memory: Up to 1.5 TB SOCAMM LPDDR5X, 1.2 TB/s bandwidth
• NVLink-C2C: 1.8 TB/s coherent bandwidth (7x PCIe Gen 6)
• Role: Orchestration, workload scheduling, KV cache routing, agentic workflow control plane
• NVIDIA's first standalone data center CPU. Jensen Huang called it "already for sure going to be a multi-billion dollar business."

Groq 3 LPU (LP30)

• Origin: NVIDIA's $20B acqui-hire of Groq (December 2025); replaces the previously announced Rubin CPX
• Generation: 3rd generation (manufactured by Samsung)
• On-Chip Memory: ~500 MB stacked SRAM per chip, ~80 TB/s bandwidth per chip
• Function: Decode-phase inference acceleration—Rubin GPUs handle prefill (processing input contexts), Groq 3 LPUs handle decode (token generation at low latency)
• When paired with Rubin NVL72, NVIDIA claims 35x higher inference throughput per megawatt for trillion-parameter models.

NVLink 6 Switch

• Bandwidth: 3.6 TB/s bidirectional per GPU (2x Blackwell)
• Per Tray: 28.8 TB/s switching bandwidth, 14.4 TFLOPS FP8 in-network compute
• Per NVL72 Rack: 9 switch trays, 260 TB/s aggregate scale-up bandwidth
• In-network compute accelerates all-to-all collective operations for MoE routing.

ConnectX-9 SuperNIC

• Throughput: 1.6 Tb/s per GPU, 8 NICs per compute tray
• Function: Scale-out connectivity (rack-to-rack) via Spectrum-X Ethernet or Quantum-X800 InfiniBand

BlueField-4 DPU

• Architecture: Dual-die, a 64-core Grace CPU + integrated ConnectX-9 NIC
• vs BF-3: 2x bandwidth, 3x memory bandwidth, 6x compute
• Function: Offloads networking, storage, encryption, and security enforcement
• Powers the new Inference Context Memory Storage Platform (CMX), extending GPU memory into NVMe storage for KV cache data.

Spectrum-6 Ethernet Switch

• Aggregate Bandwidth: 102.4 Tb/s
• Co-Packaged Optics: First NVIDIA switch with CPO, co-developed with TSMC ("COUPE")
• Efficiency: 5x power efficiency, 10x resiliency vs prior Spectrum-X generations

The Five Rack Systems

Below is a detailed breakdown of the specifications for each of the five rack systems that form the Vera Rubin platform.

Vera Rubin NVL72

• GPUs: 72 Rubin GPUs + 36 Vera CPUs
• Inference: 3.6 EFLOPS NVFP4
• Training: 2.5 EFLOPS NVFP4
• HBM4: 20.7 TB capacity, 1.6 PB/s bandwidth
• LPDDR5X: 54 TB
• Scale-Up Bandwidth: 260 TB/s (NVLink 6)
• Cooling: 100% liquid, 45°C inlet water (enables free cooling without chillers in most climates)
• Assembly: Full rack in ~2 hours (from ~2 days); compute tray swap in ~5 minutes
• Rack: ~4,000 lbs, ~1,300 chips, 1.3M components
• Power: ~190 kW (Max Q) / ~230 kW (Max P)

Groq 3 LPX Rack

• LPUs: 256
• On-Chip SRAM: 128 GB aggregate
• Memory Bandwidth: 40 PB/s
• Scale-Up Bandwidth: 640 TB/s per rack
• Designed to pair with NVL72 for disaggregated prefill/decode inference.

Vera CPU Rack

• CPUs: 256 Vera CPUs
• Designed for large-scale reinforcement learning and agentic orchestration/sandboxing.

BlueField-4 STX Rack

• AI-native storage with CMX for KV cache extension into NVMe

Spectrum-6 SPX Rack

• Silicon photonics-based networking for low-latency, resilient multi-rack connectivity

A full Vera Rubin POD scales to 40 racks, 1,152 Rubin GPUs, ~20,000 dies, and 60 exaflops.

Vera Rubin vs Blackwell

Below is a direct comparison of key performance metrics between the Vera Rubin NVL72 and the previous-generation Blackwell NVL72.

When paired with the Groq 3 LPX rack, NVIDIA claims 35x throughput per megawatt for inference, targeting what Jensen Huang framed as "premium" and "ultra" token pricing tiers for high-interactivity agentic AI.

Production Status and Availability

Vera Rubin is in full production. Rubin-based products ship H2 2026.

Cloud providers offering Vera Rubin instances: AWS, Google Cloud, Microsoft (Fairwater AI superfactories), OCI, CoreWeave, Lambda, Nebius, Nscale

Server OEMs building Vera Rubin systems: Cisco, Dell, HPE, Lenovo, Supermicro

AI labs adopting the platform: Anthropic, Cohere, Meta, Mistral AI, OpenAI, Perplexity, Runway, xAI, and others

The HGX Rubin NVL8 (8-GPU server, Intel Xeon 6 host) is also available for enterprise deployments.

Looking ahead: Rubin Ultra with the new Kyber rack (144 GPUs, 600 kW, 15 EFLOPS FP4) expected H2 2027.

Market Context

AMD's Helios rack with MI400-series GPUs also targets H2 2026, claiming 2.9 EFLOPS FP4 and 31 TB HBM4—50% more HBM4 capacity than Vera Rubin NVL72. Hyperscaler custom silicon (Google TPU, AWS Trainium, Meta MTIA) continues advancing on inference.

NVIDIA also announced the Space-1 Vera Rubin Module for orbital data centers, delivering 25x the AI compute of H100 for space-based inference. Six partners (including Axiom Space, Starcloud, and Planet Labs) are deploying on the platform, though no ship date has been set.

What Vera Rubin Means for AI/HPC Infrastructure Buyers

For Bitcoin miners building out or evaluating AI/HPC infrastructure, each new GPU generation resets the compute economics. Vera Rubin's performance-per-watt improvements and rack-scale deployment model are directly relevant to operators assessing diversification strategies. Luxor's Hardware desk sources enterprise GPUs for AI/HPC deployments. Reach out at [email protected] for pricing and availability.