Meta Platforms is accelerating its strategy to develop its own semiconductors, aiming to reduce its reliance on external suppliers like NVIDIA.

According to an internal memo seen by Reuters, Meta plans to begin mass production of its internally developed AI chip, codenamed "Iris," as early as this September. The company also intends to maintain a product cadence of releasing a new chip generation approximately every six months through 2027.

Concurrently, the social media giant has set a goal to double its total data center computing power to 14 gigawatts (GW) by 2027. This move signals Meta's push to lessen its high dependence on NVIDIA GPUs, aiming to cut AI infrastructure costs and enhance its computational autonomy.

The memo revealed that the "Iris" chip completed critical testing in just six weeks without discovering major issues. This progress is notably faster than the typical validation period of several months for similar industry chips, indicating a substantial breakthrough for Meta's in-house project.

Following the news, Meta's shares were down more than 1.5% in pre-market trading.

Focus on Cost Reduction and Reducing GPU Dependence

The "Iris" chip belongs to Meta's fourth-generation Meta Training and Inference Accelerators (MTIA) series. It will be fully custom-designed by Meta and is primarily intended to support AI training and inference tasks for products like Facebook and Instagram.

The report indicates that Meta received support from Broadcom during the chip design phase and has contracted Taiwan Semiconductor Manufacturing Company (TSMC) for production. The company has also secured its supply chain for components such as memory, flash storage, and optical interconnects. For Meta, the core objective of in-house chip development is not to completely replace GPUs but to manage the soaring costs of AI infrastructure and reduce dependence on NVIDIA and AMD.

The memo stated that deploying the latest generation of GPUs across Meta's vast infrastructure scale is "extremely difficult and has cost us time." Therefore, custom chips are set to become a crucial component of its future computing architecture.

Five-Year Investment Yields Breakthrough

Meta's journey into developing its own AI chips spans over five years, though the MTIA project has faced challenges, with several course corrections. It was once seen as a prime example of slow progress in chip development among major internet companies.

The successful completion of error validation for "Iris" in just six weeks, with no major defects found, is viewed by the industry as a significant milestone.

For data center-grade AI chips, post-silicon validation typically requires a much longer timeframe, and any major bug could necessitate a redesign or even a new tape-out. Therefore, the smooth testing of "Iris" suggests that Meta's chip development process is maturing and has cleared a key hurdle for September's planned mass production.

It is worth noting that Meta publicly disclosed the "Iris" chip and three other AI processors in March of this year. However, the details regarding the completion of testing and the target for mass production starting in September are newly revealed information.

Accelerated Development Cycle to Catch Up

Beyond the production plan, Meta is preparing to increase the pace of its chip development. The internal memo shows the company plans to maintain a rhythm of launching a new chip approximately every six months, continuing this cadence through 2027.

In contrast, leading AI chip manufacturers in the industry typically maintain product update cycles of a year or longer. A faster iteration cycle indicates Meta's desire to rapidly accumulate chip design experience through continuous upgrades. The strategy is to use its custom ASICs to complement, rather than fully replace, purchased GPUs from companies like NVIDIA in the short term.