Meta expands Nvidia deal to use millions of AI chips in data center build-out

Meta has announced a significant expansion of its collaboration with Nvidia, committing to incorporate millions of Nvidia chips into its artificial intelligence data centers. This deal, unveiled on Tuesday, features Nvidia's standalone CPUs and the advanced Vera Rubin systems, marking a landmark moment in the tech landscape. CEO Mark Zuckerberg expressed that this partnership is a crucial step in Meta's mission to provide 'personal superintelligence to everyone in the world,' a vision he first shared in July. While the financial specifics of the agreement remain undisclosed, industry expert Ben Bajarin from Creative Strategies estimated that it could be worth tens of billions of dollars. This initiative aligns with Meta's ambitious plan to allocate up to $135 billion towards AI by 2026. The extended partnership builds on a longstanding relationship, as Meta has relied on Nvidia's graphics processing units for over a decade. However, the latest agreement significantly broadens the scope of their collaboration. Notably, Meta will be the first to implement Nvidia's Grace CPUs as standalone components in its data centers, rather than integrating them with GPUs in server setups. This move is highlight as an unprecedented large-scale deployment of Grace CPUs. Bajarin noted that these CPUs are designed for inference workloads, complementing the Grace Blackwell/Vera Rubin racks. The next-generation Vera CPUs are set to be deployed by Meta in 2027, as part of a multiyear strategy that includes a commitment to invest $600 billion in U.S. data centers and associated infrastructure by 2028. Meta has plans to establish 30 data centers, with 26 located in the U.S., including two massive AI facilities currently under construction in Ohio and Louisiana. The agreement also encompasses Nvidia's Spectrum-X Ethernet switches, essential for connecting GPUs in large-scale AI environments, along with security features that will enhance AI functionalities within WhatsApp. Nevertheless, Meta is exploring other options; reports in November indicated that Meta might consider utilizing Google’s tensor processing units in its data centers by 2027. Additionally, Meta is developing its own silicon processors and has engaged with Advanced Micro Devices, which is gaining traction in the competitive AI chip market. With Nvidia's current Blackwell GPUs facing back-orders and the new Rubin GPUs recently entering production, this partnership positions Meta to secure a robust supply chain for its AI initiatives. Joint engineering teams from both companies will collaborate closely to refine and accelerate cutting-edge AI models, including Meta's new frontier model, Avocado, which aims to succeed its previous Llama AI technology. Despite recent fluctuations in Meta's stock price, including a notable decline following AI spending announcements, the company remains committed to its innovative vision in the AI domain.