Meta is preparing to begin manufacturing its next-generation artificial intelligence accelerator chip in September, according to an internal company memo outlining an ambitious roadmap to expand its AI computing infrastructure.

According to Reuters, the chip—codenamed Iris—is part of Meta’s broader Meta Training and Inference Accelerator (MTIA) program, which includes four generations of custom chips designed for AI training and inference workloads.

The company aims to strengthen the efficiency of the AI systems powering services such as Facebook and Instagram while gradually reducing its reliance on third-party hardware suppliers.

Iris Completes Early Testing

The internal memo indicates that Iris completed its validation process in just six weeks without major technical issues, marking an encouraging milestone for Meta’s in-house chip development efforts after years of investment.

The company has spent more than five years building its custom AI silicon program as demand for computing power continues to accelerate.

Broadcom and TSMC Support Production

Meta is developing the new processor in partnership with Broadcom, while TSMC will manufacture the chip.

The collaboration follows the companies’ announcement earlier this year that they were working together on 2-nanometer AI accelerators, highlighting Meta’s long-term commitment to custom semiconductor development.

Although Meta continues to purchase GPUs from NVIDIA and AMD, the new chip is designed to complement—not replace—those processors by improving efficiency across AI workloads.

The memo also notes that integrating new GPU architectures across Meta’s vast infrastructure remains a lengthy and complex process.

Aggressive AI Infrastructure Expansion

Meta first unveiled its latest AI processor lineup in March alongside three additional MTIA chips.

According to the company’s roadmap, a new generation of AI accelerators is expected approximately every six months through 2027, significantly faster than the traditional semiconductor development cycle, which often exceeds one year.

Meta plans to operate 7 gigawatts of AI computing capacity during the current year before doubling that figure to 14 gigawatts by 2027.

The company also expects to spend approximately $145 billion on AI infrastructure in 2026, representing a substantial share of the more than $700 billion analysts expect major technology companies to invest collectively in AI infrastructure.

To support that expansion, Meta has secured long-term supply agreements, while analysts at Morgan Stanley have warned that rapidly rising memory and semiconductor prices are becoming an increasing concern across the technology industry.

Long-Term MTIA Roadmap

Meta launched the MTIA initiative several years ago to reduce its dependence on expensive AI hardware supplied by NVIDIA and AMD.

The first generation debuted in 2023, focusing on recommendation systems, content ranking, and advertising algorithms. A second generation followed in 2024, delivering higher performance and improved energy efficiency before being deployed across the company’s data centers.

In 2025, Meta began testing its first internally developed chip capable of training AI models, expanding beyond inference tasks handled by previous generations.

The company’s roadmap now includes MTIA 300, MTIA 400, MTIA 450, and MTIA 500, with each generation designed to support increasingly demanding workloads, including generative AI for text, images, and video.

Vexiora Analysis

Meta’s accelerated chip development strategy reflects a broader shift in the AI industry, where competitive advantage increasingly depends not only on building powerful AI models but also on owning the computing infrastructure that powers them. Developing custom AI accelerators allows the company to improve performance, lower long-term operating costs, and reduce its dependence on external chip suppliers such as NVIDIA and AMD.

If Meta succeeds in delivering a new generation of MTIA chips every six months, it could emerge as one of the world’s leading AI semiconductor developers. As demand for AI computing continues to surge, custom silicon is expected to become a key competitive differentiator, enabling technology companies to train and deploy more advanced AI models with greater efficiency and scalability.