Meta Prepares to Produce Custom AI Chip to Double Its Computing Power

SadaNews - Meta is preparing to start production of an internally developed artificial intelligence chip in September 2026, a move that connects its rapid expansion in data centers with an effort to reduce reliance on external chip suppliers.

The chip, codenamed "Iris," according to an internal memo reviewed by Reuters, aims to increase the company's computing power from about 7 gigawatts this year to 14 gigawatts by 2027. The company has not officially announced the production schedule or the figures mentioned in the memo.

The new chip is part of the "Meta Training and Inference Accelerator" program, or "MTIA" for short, through which Meta develops custom processors to run AI workloads on its platforms.

These chips focus on tasks such as content ranking, recommendation systems, and running generative AI models, rather than fully relying on general-purpose graphics processing units purchased from suppliers like Nvidia and AMD.

In March, Meta announced that it is working on developing and deploying four new generations of "MTIA" chips over the next two years, in a faster development cycle than the traditional timelines for chip manufacturing. It stated that these generations will expand from running ranking and recommendation systems to supporting generative AI workloads.

According to the Reuters report, "Iris" passed testing within about six weeks without major issues, and the company plans to launch a new generation of custom chips approximately every six months until 2027.

Manufacturing Partnership

Meta is collaborating with Broadcom on the chip design, while Taiwan Semiconductor Manufacturing Company (TSMC) is expected to handle the manufacturing process. Meta and Broadcom expanded their partnership in April to develop several generations of custom AI accelerators. Broadcom announced that the collaboration includes a chip built using a 2-nanometer manufacturing process, along with a multi-year plan to support Meta's growing computing needs. Developing in-house chips does not mean that Meta will stop purchasing processors from other companies; custom chips typically suit specific workloads that can be optimized for the company's needs, while general-purpose processors remain essential for training large models and running a wide range of tasks. Therefore, Meta is pursuing a strategy that combines its in-house chips with processors from its partners, rather than replacing one path with another.

Massive Expansion

The goal of reaching 14 gigawatts of computing power by 2027 reveals the scale of the infrastructure the company intends to operate for developing AI models and services. Reuters notes that Meta expects to deploy about 7 gigawatts of capacity in 2026, then add a similar amount the following year. These figures represent the electrical capacity required to run data centers, chips, networking equipment, and cooling systems associated with them, not the performance of the chip by itself. This is part of a projected capital expenditure expected to range between $125 billion and $145 billion this year, according to estimates reported in follow-up reports on the company's announcement, with a large proportion of these investments directed towards data centers, chips, and AI infrastructure.

Multiple Suppliers

In addition to its collaboration with Broadcom, Meta has signed a long-term agreement with AMD to provide up to 6 gigawatts of "Instinct" processors for AI infrastructure.

Shipments for the first deployment of one gigawatt are expected to begin in the second half of 2026, using processors built on the "MI450" platform and next-generation CPUs from AMD. Meta has also partnered with Arm to develop a new class of data center CPUs, given the increasing demands of running AI exceeding traditional infrastructure capabilities in some applications.

Custom chips grant tech companies greater ability to align hardware with the software and AI models they operate, and they can help enhance energy efficiency, reduce costs, and diversify supply chains. However, Meta will still need external suppliers, whether to design some components, manufacture chips, or provide general processors. Its plan reflects a direction towards building a diverse hardware portfolio that includes in-house chips and processors from other companies, rather than relying on a single provider.