Meta pushes forward with cost-efficient AI chip development for 2027

Meta plans to deploy its next-generation self-developed AI chip, Astrid, in the first half of 2027, aiming to reduce costs and energy consumption in AI model operations. The chip, designed in collaboration with Broadcom and manufactured by TSMC, is part of Meta's broader strategy to minimize reliance on NVIDIA's AI processors. Yee Jiun Song, Meta's vice president of custom chip engineering, noted that each new chip generation involves higher technical risks but offers improved performance. Astrid, the successor to the MTIA 450, is expected to be ready in about a month and will be deployed in data centers by year-end 2027. Meta aims to deploy over 1GW of its self-developed chips within 12 months, with potential for faster deployment if AI demand remains stable. The company also mentioned that its Super Intelligent Lab is involved in chip calibration, providing insights into future AI model requirements. TSMC recently delivered 12 units of the new chip to Meta, with performance tests showing only a 2-3% deviation from simulations.