Meta to Start Mass Production of In-House AI Chips in September

Meta Platforms will begin mass production of its own artificial intelligence chips in September as it seeks to reduce reliance on Nvidia and Advanced Micro Devices and curb soaring AI infrastructure costs.

Reuters reported on July 9, citing an internal Meta memo, that the company plans to start producing the chips in September.

Meta in March unveiled four in-house AI chips, including the MTIA 400, known by the codename Iris. The MTIA 400 is part of the fourth generation of Meta's Training and Inference Accelerator project. Broadcom designed the chip, and Taiwan Semiconductor Manufacturing Co. will produce it.

The push reflects intensifying competition to secure AI semiconductors. Meta's internal memo said obtaining the latest graphics processing units had been difficult and time-consuming for large companies.

Meta is also sharply increasing investment in AI infrastructure. It plans to build 7 gigawatts of computing infrastructure this year. The company added 1 gigawatt in the first half and expects to add another 5.5 gigawatts by year-end. It plans to add 7 gigawatts next year, bringing total capacity to 14 gigawatts.

Meta's capital expenditure forecast for this year is $145 billion. Most of that will go toward AI infrastructure. The figure is about one-fifth of the $700 billion in projected capital spending by big tech companies this year.

The company is also moving to secure its supply chain. Meta said it had signed long-term supply agreements for Samsung Electronics memory chips, SanDisk flash storage and Sumitomo Electric's optical fiber equipment business.

Meta has also begun monetizing its software. The company released its AI coding model, Muse Spark 1.1, to developers through a paid API. It is the first case of Meta generating revenue from an in-house AI model, departing from its open-source Llama strategy. Pricing is $1.25 per 1 million input tokens and $4.25 per 1 million output tokens. That is more expensive than OpenAI's lower-cost models but cheaper than Anthropic's premium models.

Meta shares, which had been trading lower earlier in the session, rebounded after the release of Muse Spark 1.1. The stock closed at $631.5, up 4.7% from the previous day.

Lee Song-ryeol, Hankyung.com reporter yisr0203@hankyung.com