Meta to Reduce Reliance on Nvidia with In-House AI Chip Iris Launching Mass Production in September

Meta (META-US) is accelerating its push for AI chip self-reliance. According to an internal memo obtained by Reuters, Meta plans to begin mass production in September 2024 of its in-house AI chip codenamed 'Iris'. The company also aims to double its total computing capacity from 7GW this year to 14GW next year, supporting the rapidly increasing AI computing demands across platforms like Facebook and Instagram.

The report indicates that Iris is part of Meta's in-house Meta Training and Inference Accelerators (MTIA) program and represents one of the fourth-generation product roadmaps. Primarily targeting AI training and inference applications, the custom chip is designed to enhance the performance of Meta's AI systems while reducing reliance on external chip suppliers.

Internal documents show that the Iris chip completed testing in just six weeks with no major issues detected, signaling that Meta's in-house chip initiative—over five years in development—is yielding tangible results and laying the foundation for September's mass production launch.

Iris was designed in-house by Meta, co-developed with Broadcom (AVGO-US), and will be manufactured by TSMC (2330-TW) (TSM-US). Through this in-house chip strategy, Meta aims to reduce the high costs of AI computing, increase chip autonomy, and lessen its dependence on AI chip suppliers such as Nvidia (NVDA-US) and AMD (AMD-US).

Iris will primarily supplement, rather than fully replace, the large volumes of GPUs Meta purchases from Nvidia and AMD. Internal documents note that large-scale deployment of the latest GPUs has always been a complex and time-consuming engineering challenge for Meta, and in-house chips are expected to improve the overall efficiency of its AI infrastructure.

Meta first publicly revealed the Iris codename in March 2023, alongside three other AI processors. The company plans to maintain a cadence of launching a new AI chip approximately every six months through 2027—significantly faster than the industry's typical one-year product refresh cycle.

Beyond chip development, Meta is also expanding its AI infrastructure. The memo indicates the company will deploy 7GW of computing infrastructure this year and plans to double it to 14GW next year. AI infrastructure spending in 2024 could reach up to $145 billion, representing a significant portion of the tech giants' collective over $700 billion in AI investments.

To support large-scale data center expansion, Meta has signed multiple long-term supply agreements, including Samsung Electronics for memory, Sandisk (SNDK-US) for flash storage, and Sumitomo Electric for fiber optic equipment, ensuring stable supply of critical components. Amid tight memory supply, such long-term agreements have become essential for data center expansion and have prompted companies like Apple to raise product prices.

Fueled by the AI boom, demand for key components such as memory and AI chips continues to surge, driving rapid price increases. Morgan Stanley (MS-US) analysts note that rising chip prices have created a phenomenon dubbed 'chipflation,' impacting data center construction costs and emerging as a growing macroeconomic concern.