San Jose, California — 16 March 2026. At its annual GPU Technology Conference, NVIDIA announced that the Vera Rubin platform — a full-stack, rack-scale AI computing system integrating seven co-designed chips — has entered full production. The platform represents a generational shift in AI infrastructure, moving beyond discrete GPUs toward integrated POD-scale systems engineered for what NVIDIA CEO Jensen Huang called "the agentic AI inflection point."
The announcement positions Vera Rubin as the foundation for a new class of AI workloads: systems that perform complex, multi-step reasoning and autonomous execution rather than simple text generation. NVIDIA says the platform addresses every phase of AI — from massive-scale pretraining and post-training to test-time scaling and real-time agentic inference.
Seven Chips, Five Racks, One Supercomputer
The Vera Rubin platform is not a single GPU but a system of five purpose-built rack types, each housing different components that operate together as a single coherent machine:
| Rack Type | Key Components | Function |
|---|---|---|
| Vera Rubin NVL72 | 72 Rubin GPUs, 36 Vera CPUs, NVLink 6 Switch | Primary training and inference compute |
| Vera CPU Rack | 256 Vera CPUs | Reinforcement learning and agentic environment simulation |
| Groq 3 LPX Rack | 256 Groq 3 LPU processors | Ultra-low-latency inference acceleration |
| BlueField-4 STX Storage | BlueField-4 DPU, ConnectX-9 SuperNIC | Shared key-value cache storage across the POD |
| Spectrum-6 SPX Ethernet | Spectrum-6 Ethernet switches | Rack-to-rack fabric connectivity |
The Rubin GPU, the platform's primary compute engine, features 336 billion transistors and 288 GB of HBM4 memory per GPU. The Vera CPU is a custom ARM-based processor designed to eliminate data-feeding bottlenecks, delivering twice the efficiency and 50% faster single-threaded performance than traditional CPUs.
The Groq Integration
A notable element of the platform is the integration of the NVIDIA Groq 3 LPU — a processor originally developed by Groq, a company specializing in inference acceleration. The LPX rack houses 256 LPU processors with 128 GB of on-chip SRAM and 640 TB/s of scale-up bandwidth. At scale, a fleet of LPUs functions as a single giant processor for deterministic inference.
When deployed alongside Vera Rubin NVL72, the Rubin GPUs and LPUs jointly compute every layer of an AI model for every output token — a codesigned architecture that NVIDIA says delivers up to 35x higher inference throughput per megawatt for trillion-parameter models with million-token contexts.
Agentic AI as the Design Target
The platform's architecture reflects a shift in AI workloads. Where previous GPU generations were optimized primarily for pretraining — feeding vast datasets through models to learn patterns — Vera Rubin is designed for the full lifecycle, including post-training refinement, reinforcement learning from human feedback, test-time scaling, and agentic inference where models make autonomous decisions across multiple steps.
The Vera CPU rack addresses a specific bottleneck in agentic AI: reinforcement learning workloads require large numbers of CPU-based environments to test and validate model outputs. The 256-CPU rack provides dense, liquid-cooled infrastructure for these simulations, synchronized across the AI factory via Spectrum-X Ethernet networking.
Energy Efficiency and the DSX Platform
Power consumption has emerged as a critical constraint on AI scaling. NVIDIA addressed this with the DSX platform for Vera Rubin, developed with over 200 data center infrastructure partners. DSX Max-Q enables dynamic power provisioning across an entire AI factory, allowing deployment of 30% more AI infrastructure within a fixed-power data center. DSX Flex software enables AI factories to act as grid-flexible assets, unlocking what NVIDIA claims is 100 gigawatts of stranded grid power.
The company also released a Vera Rubin DSX AI Factory reference design — a blueprint for co-designed infrastructure that maximizes tokens per watt and overall system goodput while ensuring reliability under continuous, high-intensity workloads.
Ecosystem Adoption
Vera Rubin-based products will be available from partners starting in the second half of 2026. Cloud providers including Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure have committed to offering Vera Rubin instances, alongside NVIDIA Cloud Partners CoreWeave, Crusoe, Lambda, Nebius, Nscale, and Together AI.
System manufacturers Cisco, Dell Technologies, HPE, Lenovo, and Supermicro are expected to deliver servers based on the platform, with additional partners including Aivres, ASUS, Foxconn, GIGABYTE, Inventec, Pegatron, Quanta Cloud Technology, Wistron, and Wiwynn.
Frontier AI labs Anthropic, Meta, Mistral AI, and OpenAI have indicated plans to use Vera Rubin for training larger models and serving long-context, multimodal systems.
"NVIDIA infrastructure is the foundation that lets us keep pushing the frontier of AI," said Sam Altman, CEO of OpenAI. "With NVIDIA Vera Rubin, we'll run more powerful models and agents at massive scale and deliver faster, more reliable systems to hundreds of millions of people."
"Enterprises and developers are using Claude for increasingly complex reasoning, agentic workflows and mission-critical decisions," said Dario Amodei, CEO and cofounder of Anthropic. "NVIDIA's Vera Rubin platform gives us the compute, networking and system design to keep delivering while advancing the safety and reliability our customers depend on."
Implications for the Computing Landscape
The Vera Rubin announcement underscores several trends shaping the semiconductor and AI industries in 2026. First, the shift from discrete chips to integrated rack-scale systems reflects the reality that data movement — not raw compute — has become the primary bottleneck in AI training and inference. By co-designing compute, networking, and storage silicon, NVIDIA aims to minimize the time and energy spent moving data between components.
Second, the integration of Groq's LPU technology signals that inference — not just training — is becoming a distinct hardware market with its own optimization criteria. As AI models are deployed at scale for real-time applications, the latency and cost of generating each token becomes a critical economic variable.
Third, the focus on agentic AI as a design target reflects the industry's belief that the next wave of AI applications will involve autonomous systems that reason, plan, and act over extended time horizons — workloads that demand fundamentally different infrastructure than the chatbot interfaces that dominated the first wave of generative AI deployment.
Sources
- NVIDIA Newsroom, "NVIDIA Vera Rubin Opens Agentic AI Frontier," 16 March 2026
- NVIDIA Blog, "GTC 2026 News," March 2026
- StorageReview, "NVIDIA GTC 2026: Rubin GPUs, Groq LPUs, Vera CPUs," March 2026
- NVIDIA.com, Vera Rubin platform page
FIRAT Editorial Board
Institutional Research Desk · Foresight Institute of Research and Translation
The collective editorial and research translation board of FIRAT, synthesising peer-reviewed evidence, policy briefs, and division milestones across our seven foundational research pillars.
