Meta is accelerating its efforts to develop proprietary AI chips, with plans to deploy a new generation of chips in its data centers starting in the first half of next year. The move aims to lower the costs and power consumption associated with AI model computing while reducing reliance on Nvidia. Meanwhile, CEO Mark Zuckerberg has weighed in on the ongoing debate regarding AI safety, arguing that companies should engage independent organizations to assess model safety rather than waiting for the entire industry to hit the brakes.
Since announcing the development of its own AI chips in 2023, Meta has been testing the third-generation MTIA 450 (codenamed “Arke”) and plans to begin its deployment in the first half of next year. The design for the fourth-generation MTIA 500 (codenamed “Astrid”) is expected to be finalized in about a month, with deployment in data centers slated for late 2027 and a further expansion of usage scale.
In an interview, Yee Jiun Song, Meta’s Vice President of Engineering, stated that each successive generation of chips entails greater technical risk in exchange for superior performance—specifically, increased computational efficiency per watt of power and per dollar spent.
Meta—the parent company of Facebook, Instagram, and WhatsApp—has been aggressively expanding its AI infrastructure in recent years, with in-house chip development serving as a key component of this strategy. The company collaborates with Broadcom on chip design and utilizes TSMC for manufacturing, aiming to gradually reduce its reliance on Nvidia’s AI chips.
In terms of power consumption—a critical metric for data centers—Meta has committed to deploying in-house chips with a combined power draw exceeding 1 gigawatt (GW) within the next 12 months. Song noted that the pace of deployment is expected to accelerate further, provided there is no sudden collapse in the AI market or demand.
Meta Superintelligence Labs is also involved in chip optimization, providing data on the requirements of future AI models and inference workloads. Song stated that because Meta handles much of the engineering in-house, these chips can achieve higher efficiency than “any product currently shipping from Nvidia” when running Meta’s own AI models.
Preliminary tests indicate no significant issues with the chip design; however, several months of testing and fine-tuning are still required as the factory gradually ramps up production. All four generations of Meta’s in-house chips rely on High Bandwidth Memory (HBM) and are primarily designed for general-purpose AI inference, rather than the high-speed inference market that demands ultra-fast model responses.
Meta had originally planned to develop a chip codenamed “Olympus”—capable of handling both AI model training and inference—with a scheduled launch in 2028 or 2029. However, the project was subsequently cancelled in favor of concentrating resources on inference chips, with cost being a key factor in the decision.