Meta’s Iris AI chips head to production in September

Meta is moving custom AI silicon out of the lab. An internal memo reported by Reuters says Iris — a latest inference-focused chip in Meta’s MTIA (Meta Training and Inference Accelerator) family — cleared testing in roughly six weeks and is slated to begin production in September.

Design partners with Broadcom; manufacturing is at TSMC. Meta is also lining up Samsung for memory, Sandisk for storage, and Sumitomo Electric for fiber as it scales data-center capacity. The company still expects heavy Nvidia and AMD GPU spend for large training jobs — Iris is about owning more of ranking, recommendation, and app inference.

14 gigawatts and a $145B capex year

Meta has talked about doubling compute toward roughly 14 gigawatts by 2027 and spending as much as $145 billion on 2026 capital expenditures. A modular MTIA roadmap aims for new chips about every six months through 2027 — faster than traditional annual silicon cycles.

Google has TPUs, Amazon has Trainium, Microsoft has Maia. Iris is Meta’s version of the same hyperscaler playbook: every watt run on in-house ASICs is a watt not rented at GPU-scarce prices — even if Nvidia remains the default for frontier training.

Source: TechCrunch — Meta’s new AI chips will begin production in September.