Meta Invests in Millions of NVIDIA Chips
Meta is purchasing millions of NVIDIA chips to power its next-generation generative and vision AI systems. The massive hardware investment is part of an arms race to secure a leading position in foundation models. The partnership has reportedly contributed to a surge in NVIDIA's share price and signals the immense infrastructure demands of deploying AI at planetary scale.
- The multi-year, multi-billion dollar deal includes millions of NVIDIA's current and future chips, such as the Blackwell (B200), Vera, and Rubin series GPUs. This investment is part of Meta's broader plan to spend $600 billion on AI infrastructure by 2028, including the construction of new data centers. - This hardware is crucial for training and running Meta's large-scale AI, including the recommendation systems for Facebook and Instagram, and for developing more advanced models like Llama 3. NVIDIA CEO Jensen Huang noted that no other company deploys AI at Meta's scale for personalization and recommendation systems serving billions of users. - The deal also includes NVIDIA's standalone CPUs, the Grace and upcoming Vera processors, marking the first large-scale deployment of a data center running exclusively on Grace chips. This move signals a shift away from traditional CPU suppliers like Intel and AMD for certain workloads. - To connect these massive clusters of GPUs, Meta is also integrating NVIDIA's Spectrum-X Ethernet switches and other networking technologies to ensure high-speed communication between the chips, which is critical for large-scale model training. - The scale of the purchase is aimed at achieving Mark Zuckerberg's goal of building an open-source Artificial General Intelligence (AGI). By the end of 2024 alone, Meta's infrastructure was planned to include the equivalent of nearly 600,000 H100 GPUs. - This partnership extends to software and security, with Meta adopting NVIDIA's Confidential Computing platform to enable new AI features in WhatsApp while protecting user privacy. - While Meta is making this massive investment in NVIDIA hardware, the company is also developing its own custom silicon, the Meta Training and Inference Accelerator (MTIA), which is already being used for its ads workloads to improve efficiency. - The NVIDIA B200 "Blackwell" chip, a key part of the deal, features a dual-chip design with 208 billion transistors and offers up to 2.4 times the memory bandwidth of the previous generation H100, making it significantly faster for training and up to 15 times better for inference on the largest models.
