How NVIDIA Underpins Meta's AI Infrastructure Roadmap

Meta is deepening its multi-year partnership with NVIDIA to support the next phase of its AI-powered growth, committing to large-scale deployments of central processing units (CPUs), networking and millions of graphics processing units (GPUs) that could reshape how personal AI is delivered at global scale.

The agreement spans on-premises and cloud environments and underpins Meta's long-term AI infrastructure roadmap.

The company plans to build systems optimised for both training and inference, integrating NVIDIA Grace CPUs, Blackwell and Rubin GPUs and Spectrum-X Ethernet networking into its production architecture.

The collaboration highlights how tightly coupled compute, networking and software stacks are becoming as AI capabilities scale to serve billions of users.

Under the partnership, Meta will deploy industry-leading NVIDIA GB300-based systems and create a unified architecture across its own facilities and NVIDIA Cloud Partner environments.

The objective is to simplify operations while maximising performance and scalability for AI applications that power personalisation, recommendation systems and emerging generative features.

“No one deploys AI at Meta’s scale – integrating frontier research with industrial-scale infrastructure to power the world’s largest personalisation and recommendation systems for billions of users,” says Jensen Huang, Founder and CEO at NVIDIA.

“Through deep codesign across CPUs, GPUs, networking and software, we are bringing the full NVIDIA platform to Meta’s researchers and engineers as they build the foundation for the next AI frontier.”

Arm-based processors at scale

A core element of the agreement is the expanded deployment of ARM-based NVIDIA Grace CPUs within Meta's production applications.

“We’re excited to expand our partnership with NVIDIA to build leading-edge clusters using their Vera Rubin platform to deliver personal superintelligence to everyone in the world,” comments Mark Zuckerberg, Founder and CEO of Meta.

The companies describe this as the first large-scale Grace-only deployment, supported by joint optimisation of CPU ecosystem libraries.

The aim is to improve performance per watt with each generation, aligning with Meta's strategy to increase compute density while controlling power consumption.

Meta and NVIDIA are also collaborating on the deployment of NVIDIA Vera CPUs, with potential for large-scale rollout in 2027. For AI systems already operating at massive scale, incremental gains in performance per watt could translate into significant reductions in overall energy use.

By combining CPU optimisation with GPU acceleration, Meta is refining a heterogeneous architecture designed for large-scale AI workloads that span training frontier models and running inference for billions of daily users.

Networking for distributed intelligence

To support distributed AI clusters, Meta has adopted the NVIDIA Spectrum-X Ethernet networking platform across its infrastructure footprint.

The platform is intended to deliver predictable low-latency performance whilst maximising utilisation.

In high-performance AI environments, networking is critical to ensuring GPUs remain efficiently interconnected during model training and inference operations.

Spectrum-X will be integrated into Meta's Facebook Open Switching System platform, supporting AI-scale throughput across training and inference workloads.

The unified architecture will span on-premises facilities and partner cloud environments, providing a consistent operational model across deployments. This approach could reduce integration complexity and support faster scaling of new AI capabilities for operators.

Confidential computing for privacy

Beyond compute and networking, the partnership extends to data protection. Meta has adopted NVIDIA Confidential Computing for WhatsApp private processing, enabling AI-powered features while preserving user data confidentiality and integrity.

NVIDIA and Meta are working to expand confidential compute capabilities beyond WhatsApp to additional use cases across Meta's portfolio. For environments hosting sensitive AI workloads, confidential computing adds hardware-based security layers to protect data in use.

Engineering teams from both companies are engaged in ongoing co-design to optimise state-of-the-art AI models across Meta's core workloads.

By aligning silicon, networking and software, the companies aim to deliver higher performance and efficiency for AI capabilities deployed at global scale.

The partnership positions Meta's next-generation AI systems around a tightly-integrated NVIDIA stack, combining Grace and Vera CPUs, Blackwell and Rubin GPUs, Spectrum-X networking and confidential computing within a unified architecture designed to power personal superintelligence for billions of users.