Meta

Full briefing

Meta Platforms, Inc. (NASDAQ: META) operates Facebook, Instagram, WhatsApp, and Threads — platforms collectively serving over 3 billion daily active users. This scale makes Meta one of the world's largest AI compute consumers: every content recommendation, ad ranking, spam filter, and translation running across Meta's apps requires continuous AI inference at a scale that few other organizations can match. Meta's GPU build-out is among the most aggressive of any company. In 2024, Meta CEO Mark Zuckerberg announced plans to deploy approximately 350,000 NVIDIA H100 GPUs — a cluster larger than the US government's total AI compute capacity — for training Llama 3 and its successors. These GPUs are sourced from NVIDIA, assembled into servers by ODMs including Wiwynn and Quanta Computer, and powered/cooled by Vertiv systems. In parallel, Meta has been developing its own custom AI silicon since 2020. The MTIA (Meta Training and Inference Accelerator) v1 was deployed in Meta's data centers in 2023 for inference workloads only. MTIA v2 (2024), fabricated by TSMC on N5 (5nm), significantly improves both training and inference performance. The strategic intent is to shift recommendation model inference — Meta's highest-volume, most cost-sensitive workload — away from NVIDIA GPUs to MTIA, reducing both cost and supply chain concentration risk. Meta's open-source AI strategy (releasing Llama 2, Llama 3, and related models publicly) has a significant supply chain implication: by releasing state-of-the-art model weights freely, Meta creates a global ecosystem of Llama fine-tuning and deployment that indirectly drives demand for the full AI hardware stack — from TSMC fabrication to NVIDIA GPU sales to hyperscaler and edge deployments — further amplifying the export control stakes around advanced AI chip access.