Breaking Free from Jensen Huang's Grip? Meta's In-House AI Chip 'Iris' Set for Mass Production in September — Reuters: TSMC to Manufacture

Tech giant Meta is advancing its AI chip self-reliance strategy. According to a recent exclusive report by Reuters, Meta will begin mass production in September 2024 of its in-house AI chip codenamed 'Iris'. The chip was co-designed with Broadcom and will be manufactured by TSMC (Taiwan Semiconductor Manufacturing Company).

The report, citing an internal Meta memo obtained by Reuters, states that the mass production of Meta's self-developed AI chip will begin in September 2024 as part of the company’s broader initiative to expand its AI computing infrastructure, with a goal to scale up total computing capacity to 14GW by 2027.

The data center AI chip 'Iris' is part of Meta's fourth-generation 'Meta Training and Inference Accelerator' (MTIA) program, designed to enhance the training and inference capabilities of AI models powering platforms like Facebook and Instagram.

A Major Breakthrough for Meta's In-House Chip Development? Iris Completed Testing in Just 6 Weeks

According to the report, the internal memo reveals that the Iris chip completed testing in just six weeks, with no major issues detected. This progress signifies a clear breakthrough for Meta’s in-house AI chip initiative, which has faced criticism over slow progress since its launch more than five years ago.

The chip, custom-designed for Meta, was co-developed with Broadcom, with TSMC handling manufacturing. This strategy is expected to significantly reduce Meta’s massive AI computing costs and decrease its dependence on external AI chip suppliers such as NVIDIA and AMD.

The timeline for Iris’s debugging, testing, and mass production is being disclosed for the first time. Meta declined to comment on the report.

Will Meta Fully Replace NVIDIA and AMD? Reuters: Iris Is Just a Supplementary Solution

Additionally, the memo clarifies that Iris is not intended to replace GPUs but rather serve as a supplementary solution alongside Meta’s continued large-scale procurement of NVIDIA and AMD chips. The document notes that for a company of Meta’s scale, integrating the latest-generation GPUs remains a challenging task, increasing deployment complexity and slowing overall progress.

Meta officially unveiled Iris in March 2024, alongside three other AI processors. According to its roadmap, Meta plans to release a new AI chip approximately every six months until 2027—far faster than the industry standard of one new generation per year or longer.

Meta to Invest Another $145 Billion in Computing Power—Doubling Capacity to 14GW by 2027

The internal memo reveals that Meta plans to deploy 7GW of AI computing infrastructure in 2024 and double that to 14GW by 2027.

Meta estimates its AI infrastructure spending in 2024 could reach up to $145 billion, representing a significant portion of the total AI investment by major tech firms this year—estimated to exceed $700 billion.

Who Supplies Memory and Fiber Optics? Meta’s Long-Term Supplier Partners Revealed

The report notes that to support large-scale data center expansion, Meta has signed multi-year supply agreements with several key vendors. According to the memo, the following companies are involved: Samsung for memory, SanDisk for flash storage, and Sumitomo Electric for fiber optic equipment.

The report analyzes that amid ongoing global memory supply constraints, such long-term supply agreements have become critical safeguards for large-scale data center expansions. Recently, tech companies like Apple have raised product prices due to rising costs caused by memory shortages.

'Chipflation' Emerges as a New Market Risk? Morgan Stanley Analysts Warn of Macroeconomic Threat

The report also highlights that as global tech firms aggressively expand AI data centers, demand for key components is rising rapidly. Due to supply-demand imbalances, prices for memory and other chips have surged significantly.

Morgan Stanley analysts warn that the rapid rise in chip prices has become severe enough to pose a new macroeconomic risk. The market has even coined the term 'Chipflation' to describe the cost inflation driven by AI demand.

Related Reports: • 'DeepSeek Developing In-House AI Chips'—Reuters Reveals Biggest Loser Isn't NVIDIA: Chinese Firms on Edge • Huawei Can't Hold Out? China Reportedly to Allow NVIDIA H200 Chips—Only 'These 3 Chinese Firms' Can Buy • TSMC on Alert? Intel Partners with 'Taiwan's Second-Largest' to Develop 3nm Chips—Foreign Media: Challenging the Holy Mountain's Dominance