Meta Platforms will start producing its own AI chip, dubbed “Iris,” that it has been developing since September, an internal memo seen by Reuters revealed. The move is part of the company’s wider strategy to increase its total computing power to more than double to 14 gigawatts by next year, having come to depend on outside chip makers like Nvidia and AMD for its computing needs in the past.
The data centre chip is part of Meta’s four-generation project called Meta Training and Inference Accelerators (MTIA), and was completely built in-house to power the company’s social media giants, Facebook and Instagram. Unlike the graphics processing units (GPUs) that Meta has traditionally bought from Nvidia and AMD for its AI needs, the custom-built silicon is designed specifically for the company’s operations. According to the report, the chip went through a six-week test period without any problems, while the company has had difficulties with its previous in-house projects since their launch over five years ago.
Meta has joined forces with semiconductor giant Broadcom to help design the chip and Taiwan Semiconductor Manufacturing Company to manufacture it. The strategy is common among the hyperscale tech giants, and it aims to reduce the high computing costs that Meta can expect as its AI capabilities grow, as well as to shield the company from the pricing power of leading GPU vendors. According to the internal memo, managing to stay on top of the newest GPUs has been a “heavy lift” for a company as large as Meta, and “had taken some time.” While the Iris chip is not meant to replace the large supply of GPUs that Meta continues to purchase, it will complement them, helping to speed up both training and inference.
In March of this year, Meta has publicly announced the Iris chip—which joined three other AI processors. The company has unveiled an ambitious roadmap, with plans to release a new AI chip every six months or so, through 2027. The pace of this schedule is much more rapid than the yearly or even longer development cycles generally used by the semiconductor industry. The company is also on the way towards a huge ramp-up in physical computing infrastructure. Meta is looking to have 7 gigawatts of compute power deployed in 2026 and 14 gigawatts in 2027, the memo says. Meta has also inked long-term deals with a number of important hardware vendors, such as Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fibre-optic gear.
Google has recently joined the other major tech giants such as Amazon in making heavy investments to optimise their own chips to improve the performance and control rising infrastructure expenses. Meta earlier this year signed a new deal with Broadcom that will span several generations of MTIA chips and last until 2029, as it was agreed to provide over 1 gigawatt of computing power. Meanwhile, Meta has also announced a multi-year partnership with AMD to use up to 6 GW of AMD Instinct GPUs, adding to its range of compute sources.
Meta’s production schedule for September puts it in a sweet spot to launch Iris at scale during the heydays of massive capital expenditure on AI infrastructure. The company will spend up to US$145 billion on AI infrastructure this year, of the total excess of US$700 billion that the tech industry is projected to spend on AI this year. The custom chip proposal offers more control and cost-efficiency but also has implications on the rate of software optimisation, the fact that it would eventually have to match the performance of the most popular GPUs in the industry and the uncertainty associated with the fast semiconductor development cycles.
