Meta readies Iris chip, locks in supply for push to 14 gigawatts

Meta’s move to mass-produce its Iris AI accelerator and lock in memory, storage, and optical components signals a strategic pivot from renting compute to vertical integration. This will accelerate HBM4 adoption, co-packaged optics (CPO), and 3D-stacked storage, pressuring TSMC, Samsung, and SK Hynix to expand advanced packaging capacity ahead of schedule. Geopolitically, scaling to 14 GW of AI compute invites scrutiny under U.S. CHIPS Act export controls—especially regarding access to leading-edge nodes from Taiwan, China—forcing Meta into costly multi-source supply chains. Facing NVIDIA’s Grace Blackwell and Microsoft’s Maia, this is Meta’s bid to control the AI training stack end-to-end. Over the next 18 months, hyperscalers will race to secure integrated power-chip-optical resources, marginalizing smaller AI firms unable to shoulder custom silicon economics.