NVIDIA Infrastructure Powers Meta’s Hyperscale AI Expansion

In hyperscale AI infrastructure, large language model training, recommendation systems and real-time inference require high compute density, network throughput and energy efficiency. NVIDIA and Meta have expanded their collaboration to build large-scale AI infrastructure across on-premises and cloud environments, supporting Meta’s long-term AI roadmap.

The collaboration includes large-scale deployment of NVIDIA Grace and future Vera CPUs, millions of NVIDIA Blackwell and Rubin GPUs, integration of NVIDIA Spectrum-X Ethernet networking and adoption of NVIDIA Confidential Computing technologies.

Hyperscale deployment of CPUs and GPUs
Meta plans to build hyperscale data centers optimized for both AI training and inference. These facilities will integrate Arm-based NVIDIA Grace CPUs in production environments, representing the first large-scale deployment of Grace-only systems.

According to the companies, performance-per-watt improvements are achieved through hardware codesign and software optimization across CPU ecosystem libraries. The roadmap also includes collaboration on NVIDIA Vera CPUs, with potential large-scale deployment targeted for 2027.

On the accelerator side, Meta will deploy systems based on NVIDIA GB300 platforms, incorporating Blackwell and Rubin GPU architectures. The objective is to create a unified AI infrastructure spanning Meta’s on-premises data centers and NVIDIA Cloud Partner deployments, simplifying operations while maintaining scalability.

AI-scale networking with Spectrum-X Ethernet
To support distributed AI workloads, Meta is adopting the NVIDIA Spectrum-X Ethernet platform across its infrastructure footprint. The networking architecture is designed to deliver predictable low-latency communication and high throughput required for large model training and inference clusters.

By integrating AI-optimized Ethernet switching into Meta’s Facebook Open Switching System platform, the companies aim to improve network utilization and power efficiency at scale.

Confidential computing for privacy-preserving AI
Meta has implemented NVIDIA Confidential Computing for WhatsApp private processing. This architecture enables AI-driven features while maintaining confidentiality and integrity of user data during processing.

The collaboration includes expanding confidential computing capabilities beyond WhatsApp to additional Meta services, supporting privacy-enhanced AI workloads across the company’s application portfolio.

Codesign for next-generation AI models
Engineering teams from both companies are engaged in joint hardware-software codesign to optimize AI model performance across Meta’s production workloads. This includes tuning compute, memory and networking stacks to support large-scale personalization and recommendation systems serving billions of users.

By combining CPU, GPU, networking and security technologies within a unified architecture, the initiative targets improvements in compute efficiency, system scalability and operational management for hyperscale AI infrastructure.

www.nvidia.com