NVIDIA Infrastructure Powers Meta’s Hyperscale AI Expansion

[Cengiz Özemli]

## NVIDIA Infrastructure Powers Meta’s Hyperscale AI Expansion

Meta is leveraging NVIDIA Grace and Vera CPUs, Blackwell and Rubin GPUs, Spectrum-X networking technology, and Confidential Computing solutions to scale AI training and inference processes across its global data centers.

In hyperscale AI infrastructure, large language model training, recommendation systems, and real-time inference demand high computational intensity, network bandwidth, and energy efficiency. The collaboration between NVIDIA and Meta aims to build large-scale AI infrastructure in both on-premise and cloud environments, supporting Meta’s long-term AI plans.

### CPU and GPU Hyperscale Deployment

Meta plans to establish hyperscale data centers optimized for both AI training and inference. These centers will integrate Arm-based NVIDIA Grace CPUs into production environments, marking the first large-scale deployment of Grace-based systems.

Through co-design of hardware and software optimization, CPU performance per watt is being increased, and collaboration on NVIDIA Vera CPUs continues, with large-scale deployments targeted to begin in 2027.

On the GPU side, Meta will utilize Blackwell and Rubin GPU architectures based on NVIDIA GB300 platforms. The goal is to create a unified and scalable AI infrastructure between Meta’s on-premise data centers and NVIDIA Cloud Partners, streamlining operations.

### AI-Scale Networking with Spectrum-X Ethernet

Meta is integrating the NVIDIA Spectrum-X Ethernet platform into its infrastructure to support distributed AI workloads. This network architecture provides low-latency, high-bandwidth for large model training and inference, offering performance efficiency.

AI-optimized Ethernet switching technology is being integrated into Meta’s Facebook Open Switching System platform, increasing network efficiency and energy savings.

### Privacy-Focused AI with Confidential Computing

Meta has deployed NVIDIA Confidential Computing technology for WhatsApp. This structure ensures the secure use of AI-based features while protecting the privacy and integrity of user data.

In the future, these confidential computing methods are planned to be extended to other Meta services beyond WhatsApp, thereby supporting privacy-focused AI workloads across the company.

### Co-Design for Generative AI Models

Meta and NVIDIA engineering teams are working to optimize AI model performance through co-design of hardware and software. This aligns compute power, memory, and networking infrastructure with large-scale personalization and recommendation systems.

By integrating CPU, GPU, networking, and security technologies into a single architecture, the aim is to achieve advancements in computational efficiency, system scalability, and operational management within hyperscale AI infrastructure.