An internal memo revealed that Meta plans to start manufacturing its new AI chip as of next September, as part of an ambitious plan to increase its total computing capacity to 14 gigawatts by 2027.

According to Reuters, the chip, codenamed Iris, is part of a project spanning four generations of training and inference accelerators within the MTIA chip family that Meta is developing internally.

The company's strategy relies on designing custom chips to enhance the efficiency of the AI systems that power its services and platforms, foremost among them Facebook and Instagram.

The memo showed that the chip's testing phase took only six weeks, without detecting any major issues, which is a positive indicator for the company's chip development efforts, after the project faced several challenges since its launch more than five years ago.

Expanded Partnerships

Meta is collaborating on developing the chip with Broadcom, with whom it announced in April the start of work on producing AI accelerators using 2-nanometer technology, while Taiwan Semiconductor Manufacturing Company (TSMC) handles the manufacturing processes.

This approach is expected to help reduce high computing costs, as well as decrease the company's reliance on major chip suppliers such as Nvidia and AMD.

Despite developing Iris internally, the chip is not intended to replace the graphics processing units that the company purchases from Nvidia and AMD, but rather to support them and enhance their capabilities in running AI applications.

The memo indicates that integrating the latest graphics processing units within a company the size of Meta requires significant efforts and long periods of time.

Meta had revealed its processor for the first time in March, along with three other AI processors.

The company plans to launch a new chip approximately every six months until 2027, a faster pace than usual in the chip industry, where launch cycles typically take a year or more.

The memo explains that Meta aims to operate computing infrastructure with a capacity of 7 gigawatts during the current year, and to double that number to 14 gigawatts next year.

The company also expects to spend about $145 billion on AI infrastructure during 2026, which represents a large part of the total expected spending by major technology companies on this sector, which exceeds $700 billion.

The memo added that Meta has signed long-term supply agreements to support its expansion in computing infrastructure, at a time when analysts at Morgan Stanley warned of a rapid and significant rise in memory and chip prices, noting that 'chip price inflation' has become a growing macroeconomic concern.

A Long-Term Project

It is worth noting that Meta began working on developing its own AI chips several years ago under the Meta Training and Inference Accelerator project, known as MTIA, in an attempt to reduce its reliance on high-cost chips from companies such as Nvidia and AMD.

The company launched the first generation of those chips in 2023 to run recommendation systems, content ranking, and ads across its platforms, before revealing in 2024 the second generation, which provided higher performance and greater energy efficiency, and was widely deployed in the company's data centers.

The project received a major boost during 2025 when Meta began testing the first internally designed chip for training AI models, after previous generations focused only on inference tasks and running models.

In March, Meta unveiled a new roadmap that includes four consecutive generations of MTIA chips, including the MTIA 300, MTIA 400, MTIA 450, and MTIA 500 processors, with a plan to launch a new generation every six months or so until the end of next year.

These chips aim to expand from running recommendation and advertising systems to supporting generative AI applications, including text, image, and video generation, in a step that reflects the company's ambition to build an independent computing infrastructure that supports its plans for advanced AI and superintelligence in the future.