Meta is preparing to begin production of its latest AI-focused chips in September as it looks to reduce its reliance on expensive GPUs during an ongoing global component shortage, according to a Reuters report citing an internal company memo.
The memo revealed that at least one of the new chips successfully completed testing in just six weeks, marking a significant milestone in Meta’s in-house semiconductor development. The company collaborated with Broadcom on the chip design, while Taiwan Semiconductor Manufacturing Company (TSMC) will handle production. Other key suppliers include Samsung for memory, Sandisk for storage, and Sumitomo Electric for fiber-optic components.
The new processors are part of Meta’s Meta Training and Inference Accelerator (MTIA) program. Introduced in March, the latest MTIA lineup consists of four chips, with some already being deployed and others scheduled for release later this year and into next year. Meta says the chips are built using a modular architecture, allowing future generations to evolve more quickly as AI workloads continue to change.
According to the company, each new MTIA generation builds upon previous designs by using modular chiplets, integrating the latest AI hardware innovations, and shortening development cycles to keep pace with the rapid evolution of artificial intelligence.
By expanding its custom silicon efforts, Meta aims to reduce the cost of purchasing GPUs from industry leaders such as Nvidia and AMD. However, the company is expected to continue investing heavily in third-party GPUs alongside its own chips. The MTIA processors will be used to train recommendation and ranking models, support broader AI workloads, and power inference across Meta’s applications. Meta has been developing its own AI chips since 2023.
The chip initiative is part of Meta’s much larger AI investment strategy. In April, the company projected capital expenditures of between $125 billion and $145 billion for the year, with a substantial portion dedicated to AI infrastructure.
To support its expanding AI ambitions, Meta has also been investing aggressively in data centers and power infrastructure worldwide. The company is committing tens of billions of dollars to secure the computing capacity needed for training and deploying its latest Muse Spark family of AI models. According to the Reuters report, Meta plans to deploy 7 gigawatts of computing capacity this year and double that figure next year, underscoring the scale of its long-term AI strategy.
Meta has also expanded its AI infrastructure through major partnerships. Last year, the company signed an agreement with Arm to strengthen computing power for its recommendation systems, secured a multibillion-dollar deal with AMD for its Instinct GPUs, and partnered with Amazon in another multibillion-dollar agreement to use the cloud provider’s custom-built CPUs for AI workloads.
Meta is far from alone in its effort to reduce dependence on Nvidia. OpenAI recently introduced a custom inference chip developed in collaboration with Broadcom, while Anthropic is reportedly exploring the development of its own AI processors with Samsung. Meanwhile, Amazon and Google continue to invest heavily in designing proprietary chips for AI training and inference, as a growing number of startups enter the market to meet surging demand for AI computing power.
Meta declined to comment on the report.
Source: techcrunch Edited By Bernie
