According to an internal Meta memo, Meta Platform Company plans to launch mass production of self-developed artificial intelligence chips in September this year, with the supporting goal of increasing the overall computing power scale to 14 gigawatts in 2027. This data center chip, codenamed “Iris”, is self-developed by MetaMTIA (Meta Training and Inference Acceleration)Fourth generation chip research and development project. Meta plans to rely on self-developed customized chips to optimize various AI businesses supporting the two major social platforms of Facebook and Instagram.
The memo disclosed that the chip only took six weeks to complete the full-process defect testing, and no major technical problems were found. The speed of research and development has greatly exceeded expectations, marking an inflection point for Meta's self-developed chip project; the progress of the project has been unsatisfactory for more than five years since its launch.
The Iris chip is completely customized for Meta's own business needs. Broadcom participates in joint design and TSMC is responsible for foundry production. This self-developed solution is expected to significantly reduce the company's huge computing power procurement costs and reduce its dependence on external chip manufacturers such as Nvidia and AMD.
Chip defect testing has been completed and plans for mass production in September have never been disclosed before. Meta officials declined to comment for this story.
This self-developed chip will be used as a supplement and paired with Meta’s large-volume graphics processors (GPUs) purchased from Nvidia and AMD to jointly carry various AI computing tasks.
The memo bluntly stated that for large technology companies like Meta to implement the latest generation of GPUs on a large scale, "it is extremely difficult to implement and consumes a lot of time and cost for the company."
Meta announced the name of the Iris chip technology in March this year, and three other AI processors were also unveiled at the same time. The company plans to launch a new self-developed chip every six months on average until 2027; the industry's normal pace is usually one year or more to update a new generation of AI chips.
Computing power construction plan: 7 GW in 2026, doubled to 14 GW in 2027
The memorandum shows that Meta will implement a computing power infrastructure with a total scale of 7 GW this year; in 2027, the scale of computing power will be doubled to 14 GW.
The company expects to invest up to US$1.45 trillion in AI infrastructure this year, accounting for a significant portion of global technology giants’ total AI capital expenditures of more than US$700 billion.
In order to ensure the expansion of computing power, Meta has signed a multi-year long-term supply agreement with partners including Samsung Electronics (memory chips), SanDisk (flash storage), and Sumitomo Electric (fiber optic communication equipment).
The global supply of memory chips is currently in short supply, and companies such as Apple have raised product prices due to supply shortages. This type of long-term price-locked supply agreement has become a key guarantee for major companies to achieve their data center expansion goals.