Meta plans to deploy its own AI chip MTIA next year to cut AI costs

Meta plans to deploy a new generation of its own AI chips in data centers in the first half of next year, aiming to cut the cost and energy use of running AI models. The company is now testing its third-generation product, MTIA 450, and the next version, MTIA 500, is expected to finish design in about a month and enter data centers by the end of 2027, with a larger rollout planned.

The chip is designed with Broadcom and made by TSMC, partly to reduce reliance on Nvidia's leading AI processors. Meta plans to deploy more than 1GW of its own chips within 12 months. Engineering VP Yee Jiun Song said the chips can run Meta's own AI models more efficiently than anything Nvidia currently ships.

Meta has canceled its Olympus project, which was meant for both training and inference, and is focusing on inference chips instead. Cost is a major reason: handling both would raise chip costs by 30%, which is unacceptable at multi-gigawatt scale.