Particle.news

Meta to Begin Producing Its Iris AI Chip in September 2026

The company says the custom processor will lower inference costs, reduce reliance on third-party GPUs, and support a push to 14 gigawatts of AI compute by 2027.

Overview

  • Reporting based on internal memos shows Meta expects Iris to clear final tests and enter production in September 2026 to run high-volume inference across Facebook, Instagram, and WhatsApp.
  • Iris is the first in a multi‑generation MTIA (Meta Training and Inference Accelerators) roadmap that Meta plans to refresh roughly every six months to handle ranking, recommendation, and generative workloads.
  • Meta has enlisted Broadcom for chip design and TSMC to manufacture Iris while signing long‑term supply agreements for high‑bandwidth memory, flash storage, and fiber optics to back volume deployment.
  • The chip program is tied to an aggressive infrastructure plan that seeks to double AI capacity from about 7 gigawatts in 2026 to roughly 14 gigawatts in 2027 and sits alongside $125–$145 billion in 2026 capital spending.
  • Execution risks remain significant because of the fast iteration cadence, huge power and supply‑chain demands, and Meta’s continued use of merchant GPUs from Nvidia and AMD as a backup during rollout.