Meta Readies Custom Iris AI Chip for September Production
Meta Platforms plans to begin manufacturing its custom AI chip, codenamed Iris, in September 2026. The chip is part of a four-generation roadmap designed to reduce Meta's dependence on Nvidia and AMD GPUs while supporting AI features across Facebook and Instagram.
Meta Platforms plans to start manufacturing its custom data-center AI chip in September 2026. The chip, codenamed "Iris," is part of a four-generation roadmap the company calls MTIA, short for Meta Training and Inference Accelerator.
Internal documents reviewed by The Verge show Meta aims to scale its computing infrastructure to 14 gigawatts by 2027, up from seven gigawatts targeted for 2026. The Iris chip was developed with Broadcom and will be manufactured by TSMC. Testing completed in six weeks, which Meta described as ahead of schedule.
The move is part of Meta's push to reduce its dependence on Nvidia and AMD graphics processing units. The company projects up to $145 billion in AI infrastructure spending this year. Custom chips allow Meta to optimize hardware specifically for its workloads, potentially cutting costs and improving performance compared to general-purpose GPUs.
Iris will support AI features across Facebook, Instagram, and WhatsApp. Meta has been integrating AI tools into its platforms throughout 2026, including multimodal features that can process text, images, and video together.
Energy supply and TSMC production capacity are the two main constraints on Meta's timeline. The company is competing with other large technology firms for access to advanced chip manufacturing slots and power infrastructure for its data centers.
The broader chip race is intensifying across the industry. Intel committed 5 billion euros to expand manufacturing in Ireland. India commissioned its third semiconductor plant. Chinese startup DeepSeek is developing its own inference chip. Anthropic is exploring custom chip manufacturing with Samsung.
Analysts say Meta's vertical integration strategy mirrors moves by Google, which has developed its own Tensor Processing Units, and Amazon, which builds custom chips for its AWS cloud platform. The goal in each case is to reduce reliance on third-party suppliers and gain more control over the cost and performance of AI workloads.
