The company’s data center chip, codenamed “Iris,” is part of a multi-generation initiative for Meta Training and Inference Accelerators (MTIA) that will be designed in-house. The goal is to utilize custom silicon to enhance the AI that supports its Facebook and Instagram platforms.
The chip’s testing only took six weeks, with no significant issues reported, indicating positive progress for an in-house project that has struggled since its inception over five years ago.
Meta has specifically tailored the chip to meet its requirements and is collaborating with Broadcom for its design and Taiwan Semiconductor Manufacturing Co for its production. This strategy is expected to help lower the company’s substantial computing costs and provide greater autonomy from chip suppliers like Nvidia and Advanced Micro Devices.
The completion of bug testing and the timeline for production have not been disclosed previously. Meta chose not to comment.
The chip aims to supplement the large volume of graphics processing units (GPUs) that Meta acquires from Nvidia and AMD for its AI applications.
However, implementing the latest GPUs within a company of Meta’s scale “has been a heavy lift, and it has cost us time,” the memo revealed.
Meta introduced Iris under its technical designation in March, along with three additional AI processors. The company plans to release a new chip approximately every six months through 2027, in contrast to the usual annual release cycle for AI chips.
SEVEN GIGAWATTS OF COMPUTING IN 2026
This year, Meta aims to deploy seven gigawatts of computing infrastructure, with plans to double that amount in 2027, according to the memo.
The company anticipates spending up to $145 billion on AI infrastructure this year, a sizable part of the more than $700 billion that Big Tech is projected to invest in the technology.
To support the expansion of its computing infrastructure, Meta has established long-term, multi-year supply agreements, as outlined in the memo. These agreements involve partnerships with Samsung Electronics for memory chips, Sandisk for flash storage, and Sumitomo Electric for fiber-optic equipment.
Such long-term agreements have become essential for achieving data center expansion goals amid a memory chip shortage that has led companies like Apple to increase prices.
Sandisk declined to provide a comment, while Samsung Electronics and Sumitomo Electric did not respond to inquiries for comment.
Components such as memory and AI chips have seen a surge in demand as technology firms rush to expand data centers to meet AI’s increasing need for computing power.
Prices for memory and other chips have risen sharply enough that “chipflation” is now a macroeconomic concern, according to analysts at Morgan Stanley.
