Meta’s AI Chips Poised for Production Launch in September

To cut GPU costs amid a significant component shortage, Meta plans to start producing its latest AI-focused chips in September, according to a Reuters report citing an internal memo.

The memo revealed that at least one chip successfully completed its testing phase in around six weeks. While Meta is working with Broadcom on the chip design, Taiwan’s TSMC will manage the manufacturing. The company is also sourcing RAM from Samsung, storage from Sandisk, and fiber optic components from Sumitomo Electric, as noted in the report.

In March, Meta disclosed information about four new chips being developed under its Meta Training and Inference Accelerator (MTIA) program, with some expected to be deployed this year or next. The company is employing a modular design strategy for these chips, anticipating that their needs will evolve as AI technology progresses by the time they go into production.

“Each generation of MTIA builds upon the last, utilizing modular chiplets that incorporate the latest advancements in AI workloads and hardware technologies, and deploying them on a faster timeline,” the company announced at the time.

These chips are expected to help the company reduce expenses on GPUs from manufacturers like Nvidia and AMD; however, significant spending with these vendors is still projected, as reported by Reuters. Meta intends to use the MTIA chips for training models that support ranking and recommendation algorithms, as well as for broader AI workloads and inference in its applications. The social media giant has been developing its own AI chips since 2023.

Meta has been heavily investing in obtaining the necessary computing power for its various AI initiatives. In April, the company announced it expects capital expenditures between $125 billion and $145 billion this year, with much of this funding directed toward AI projects.

The company has been signing agreements for data centers and power globally, spending tens of billions to secure the computing resources needed to train and launch its new Muse Spark series of AI models. It aims to deploy 7 gigawatts of computing capacity this year, with plans to double that in the following year, as cited by Reuters in the memo.

Moreover, Meta signed a deal with ARM last year to secure computing resources for its recommendation systems, alongside a multi-billion-dollar agreement with AMD for its Instinct GPUs, and another significant deal with Amazon to leverage the cloud giant’s proprietary CPUs for AI-related applications.

Meta is not the only company aiming to reduce its financial reliance on Nvidia. Last month, OpenAI unveiled an inference processor it is co-developing with Broadcom, while Anthropic is reportedly looking into creating its own chips with Samsung. Both Amazon and Google manufacture their own chips for AI training and inference, with numerous startups emerging to meet the soaring demand.

Meta declined to comment.