Meta Platforms, Inc., the owner of Facebook, has begun testing its first in-house chip designed for training artificial intelligence systems, marking a pivotal development as the company aims to create more of its own silicon and decrease dependence on external suppliers like Nvidia, according to two sources cited by Reuters.
Meta begins testing its own AI training chip amid Nvidia reliance
The social media giant has initiated a small deployment of this new chip and plans to expand its production for widespread use if initial tests prove successful. This move is part of Meta’s long-term strategy to reduce operational costs, particularly as it invests heavily in AI tools for growth. The company has forecasted total expenses for 2025 to be between $114 billion and $119 billion, which includes up to $65 billion in capital expenditures primarily driven by AI infrastructure investments.
One source indicated that Meta’s new training chip functions as a dedicated accelerator, designed specifically for AI tasks, which can enhance its power efficiency compared to standard graphics processing units that are typically employed for AI workloads. Meta is partnering with Taiwan-based chip manufacturer TSMC (2330.TW) for the production of this chip.
The testing phase commenced after Meta completed its first “tape-out” of the chip—a critical step in silicon design that involves submitting an initial design to a chip fabrication facility. This phase typically costs tens of millions of dollars and can take three to six months, with no guarantee of success. Should the testing fail, Meta would need to diagnose the issue and initiate another tape-out.
The chip represents the latest iteration in Meta’s Meta Training and Inference Accelerator (MTIA) series, which has experienced setbacks, including the abandonment of a previous chip during a similar stage of development. However, last year Meta began utilizing an MTIA chip for inference processes that govern content recommendations on Facebook and Instagram.
How Singapore became a hotspot for smuggled Nvidia AI chips
Company executives have expressed intentions to implement their proprietary chips by 2026 for training, which involves processing vast amounts of data to “teach” the AI systems. Initially, the training chip will focus on recommendation systems, with future applications in generative AI products, such as the chatbot Meta AI. Chris Cox, Chief Product Officer, stated that the company is exploring training for recommender systems and the potential for training and inference in generative AI, characterizing the chip development process as a “walk, crawl, run situation.” He noted that the first-generation inference chip is considered a significant success.
Previously, Meta abandoned an in-house custom inference chip after an unsuccessful small-scale test, leading the company to order billions of dollars’ worth of Nvidia GPUs in 2022. Since then, Meta has remained one of Nvidia’s largest customers, acquiring an extensive array of GPUs for training its models, including those used for recommendation algorithms and ad systems, as well as its Llama foundation model series. These GPUs also facilitate inference for over 3 billion daily users of its applications.
This year has seen scrutiny of the value of these GPUs as AI researchers question the efficacy of merely scaling large language models by increasing data and computational power. Concerns intensified after the January release of competitively priced models by the Chinese startup DeepSeek, which emphasize computational efficiency through enhanced inference use over traditional models.
Nvidia experienced a significant drop, losing up to a fifth of its value during a sell-off in AI stocks spurred by DeepSeek’s innovations, although shares later recovered as investors predicted Nvidia’s chips would remain essential for training and inference, despite subsequent declines tied to broader trade issues.
Meta’s latest chip development comes amid broader industry trends, with major players like OpenAI reportedly finalizing in-house designs, collaborating with companies like Broadcom and TSMC. Sources indicate that TSMC is responsible for producing the test batches for Meta, while other industry reports suggest Meta may be working with Broadcom on the tape-out process for its new AI training accelerator.
The development of the MTIA series has been ongoing for years, having experienced initial challenges that led to previous chip designs being discarded. Last year, Meta began employing an MTIA chip for user interaction processes tied to its AI systems. The urgency for custom silicon solutions is evident as the company aims to have these systems operational for AI training by next year, though it remains unclear if the new chip will utilize an open-source RISC-V architecture like previous MTIA hardware.
Featured image credit: Dima Solomin/Unsplash
