Meta plans to deploy a new generation of custom AI chips in its data centers in the first half of next year. The company aims to reduce operational costs and energy consumption while running AI models. Currently, Meta is testing the third-generation MTIA 450 chip, codenamed Arke.
The next-generation chip, MTIA 500 codenamed Astrid, is expected to complete its design phase in about a month, with a data center rollout scheduled for late 2027. Meta intends to deploy Astrid on a larger scale than its predecessors. The chips are designed in collaboration with Broadcom and manufactured by TSMC.
Meta aims to deploy over 1 GW of custom chips within 12 months, with plans to accelerate this pace depending on market demand. On September 1, TSMC delivered 12 new chips to Meta, which showed a performance deviation of only 2% to 3% compared to simulations. These chips successfully ran Meta's own models, as well as those from DeepSeek and Alibaba.
The company has shifted its focus toward specialized inference chips, canceling the Olympus project that was intended to handle both training and inference. According to engineering VP Yee Jiun Song, combining these tasks would have increased chip costs by 30%, which the company deemed unacceptable at the gigawatt scale.
Future development phases will prioritize increasing chip speed and throughput, alongside the integration of optical fiber technology. Meta intends to continue developing custom hardware that can compete with existing third-party suppliers, positioning these chips as the primary solution for general-purpose inference tasks.
