March 11 (Reuters) - Social media company Meta(which owns Facebook, Instagram, and WhatsApp) is testing its first homegrown artificial intelligence system for trainingchip. According to two sources with knowledge of theThis move marks a significant step in Meta's efforts to reduce the need for external suppliers such asNvidiaThe company has also taken a key step towards designing more customized chips.
Meta has begun deploying the chip on a small scale and plans to increase production for large-scale use if testing goes well, according to the report.Meta's push to develop an in-house chip is part of its long-term plan to...The aim is to reduce the huge infrastructure costs.
Meta's total spending in 2025 is forecast to be between $114 billion and $119 billion (note: currently around Rs. 827,736 crore to Rs. 864,041 crore), including up to $65 billion in capital expenditures, much of which will be spent on AI infrastructure development. The source said.This new training chip from Meta is a dedicated gas pedal that specializes in AI-related tasksThis gives it an energy-efficiency advantage over the integrated graphics processing units (GPUs) typically used for AI workloads.
also,Meta is working with Taiwanese Chinese chipmaker TSMC to produce the chipThe test deployment was initiated after Meta completed its first "flow" of chips. The test deployment was initiated after Meta completed the first "flow" of the chip. Flow-through is an important milestone in silicon development that involves getting the initial design into the chip factory for production. Typically, flow-through costs tens of millions of dollars, takes about three to six months, and there is no guarantee that testing will be successful. If it fails, Meta will have to diagnose the problem and re-run the flow.
While Meta's plans for a homegrown chip have had their ups and downs over the past few years, even abandoning a chip in a similar stage of development at one point, the company is continuing to move forward. Last year, Meta began using a homegrown inference chip in its recommender system to run the AI systems in Facebook and Instagram news pushes, and Meta executives said they plan to start using the homegrown chip in 2026 to "train" the AI systems to perform computationally-intensive processes by feeding them large amounts of data. Meta executives said they plan to start using their own chips for training in 2026, which involves feeding the AI systems large amounts of data to "train" them to perform computationally intensive tasks.
Speaking at last week's Morgan Stanley Technology, Media & Telecom conference, Meta chief product officer Chris Cox said, "We're looking at how to train for recommender systems and how to think incrementally about training and reasoning for generative AI." He described Meta's chip development as a process of "crawling to walking to running," but noted that the first generation of inference chips for recommender systems was a "huge success."
However, Meta had previously suspended projects after a homegrown inference chip failed in a similar small-scale test deployment, and instead ordered billions of dollars' worth of GPUs from NVIDIA in 2022.Since then, Meta has remained one of NVIDIA's largest customers, purchasing a large number of GPUs to train its models, including recommender systems, advertising systems, and its Llama base model family. These chips also provide inference services to more than 3 billion users of Meta applications every day.
The value of these GPUs has been questioned this year as AI researchers have expressed skepticism about the potential to "scale" large language models with ever-increasing amounts of data and computational power. This skepticism was reinforced at the end of January by the introduction of new low-cost models from Chinese startup DeepSeek, which optimize efficiency by relying more on inference than computational power. As a result of DeepSeek, the global AI stock market has seen significant volatility, with NVIDIA's shares falling by a fifth at one point, and although most of those losses have since been recouped, they have recently fallen again due to trade concerns and other factors.
