Zuckerberg and Jensen Huang sign a massive agreement Meta will deploy millions of NVIDIA chips

Interface News has learned that on February 17th local time, NVIDIA and Meta announced a multi-year, cross-generational strategic partnership, covering local deployment, cloud, and AI infrastructure.

Specifically, Meta will build large-scale data centers for AI training and inference to support the company’s long-term AI infrastructure roadmap. This collaboration will enable the scaled deployment of NVIDIA CPUs and millions of Blackwell and Rubin architecture GPUs, while integrating NVIDIA Spectrum-X Ethernet switches into Meta’s Facebook Open Switching System platform.

NVIDIA founder and CEO Jensen Huang stated, “No other company in the world can match Meta’s AI deployment scale. Through collaborative design across CPUs, GPUs, networking, and software layers, NVIDIA’s full-stack platform will empower Meta’s R&D teams.”

Meta founder and CEO Mark Zuckerberg said, “I am pleased to deepen our cooperation with NVIDIA and build top-tier computing clusters using their Vera Rubin platform.”

Notably, Meta has become the first company globally to independently deploy NVIDIA Grace CPUs at scale. Interface News learned that the Grace CPU is NVIDIA’s first flagship CPU project based on the Arm architecture, designed for data centers, AI, and high-performance computing (HPC). Named after computer pioneer Grace Hopper, it was first released in April 2021. Its core purpose is to address the power and performance conflicts in data centers for the AI era using high-efficiency Arm architecture.

The two companies will also collaborate on deploying NVIDIA Vera CPUs, which are expected to achieve large-scale application by 2027, further expanding Meta’s high-efficiency AI computing capabilities.

Additionally, both companies disclosed that Meta will deploy systems based on NVIDIA GB300, creating a unified architecture covering local data centers and NVIDIA cloud partner deployment environments. This aims to simplify operations while maximizing computing power, performance, and scalability. Meta will adopt NVIDIA Spectrum-X Ethernet network platforms across its entire infrastructure to provide AI-grade networking capabilities.

This cooperation is not merely about chip procurement but involves full-chain integration, leading to speculation about whether Meta will abandon in-house chip development. Last March, there were reports that Meta had begun testing its first autonomous AI training chip, in collaboration with TSMC.

Meta has not provided many updates on its in-house chip development. However, based on the cooperation with NVIDIA, it is likely that Meta will position its self-developed chips as supplementary for specific scenarios, with the main computing power shifting to mature NVIDIA solutions. This approach aims to shorten deployment cycles, reduce technical risks, and quickly support the iteration of personal superintelligence and large models, while maintaining long-term technological reserves and ensuring rapid AI business development.

Currently, major tech giants generally adopt a “self-developed + procurement” model for AI chips. For example, Google developed its own TPU chips but also heavily procured NVIDIA GPUs, while Amazon has developed its own Trainium chips and also procures chips from NVIDIA and AMD.

As of the close of U.S. markets on February 17th, NVIDIA closed at $184.97, up 1.18%, with a total market value of $4.49 trillion, remaining the most valuable company in the U.S. stock market. Meta closed at $639.29, down slightly by 0.08%, with a market value of $1.62 trillion, ranking ninth in market capitalization among U.S. stocks.

(Article source: Interface News)