Meta Expands NVIDIA AI Chip Alliance
Meta and NVIDIA have expanded their multiyear AI partnership, with Meta committing to deploy millions of NVIDIA's upcoming Blackwell and Rubin-generation GPUs and CPUs. The hardware will be used to power Meta's data centers for both large-scale AI model training and inference services. This significant investment underscores the intense competition for advanced AI compute infrastructure.
- Meta's capital expenditure on AI infrastructure is projected to reach between $115 billion and $135 billion in 2026, with a commitment to spend at least $600 billion by 2028 on U.S. data centers and related infrastructure. - The Blackwell B200 GPU, built on a custom TSMC 4NP process, features 208 billion transistors and is a dual-die chip, offering up to 5 times the AI performance of its Hopper predecessor. The follow-on "Rubin" GPU architecture is expected to feature 336 billion transistors and utilize next-generation HBM4 memory. - The collaboration extends beyond GPUs to include NVIDIA's CPUs and networking hardware. Meta will be the first company to deploy NVIDIA's Grace CPUs at scale on their own and will adopt the next-generation "Vera" CPUs starting in 2027. - To connect the massive GPU clusters, Meta will integrate NVIDIA's Spectrum-X Ethernet networking platform, which is designed to provide low-latency performance optimized for large-scale AI workloads. - This major investment in NVIDIA hardware comes as Meta continues to develop its own custom silicon, the Meta Training and Inference Accelerator (MTIA), in an effort to reduce long-term reliance on a single supplier for some of its AI workloads. - On the software and security side, Meta will use NVIDIA's Confidential Computing technology to enable new AI-powered features within WhatsApp while maintaining user data privacy and integrity. - The primary competitors in the data center AI chip market include AMD, with its Instinct MI300 series, and Intel, with its Gaudi line of accelerators. Other large tech companies, including Google and Amazon, are also investing heavily in developing their own custom AI chips.
