Meta (META.US) doubles down on self-developed AI chips! Deployment in data centers in the first half of next year, aiming to reduce inference costs and energy consumption.

Meta Platforms (META.US) plans to deploy its new generation of self-developed AI chips to data centers in the first half of 2027, hoping to reduce the energy consumption and cost of running AI models through custom chips. At the same time, the company has already committed a deployment scale of more than 1 gigawatt for the related chips, and said that if AI demand continues to remain strong, the pace of subsequent deployment will further accelerate. Meta engineering vice president Yee Jiun Song said in an interview that the company's third-generation self-developed AI processor MTIA 450 is currently in the testing phase, with the chip codenamed "Arke"; the next-generation MTIA 500, codenamed "Astrid," is expected to complete design in about a month and is planned to enter data centers before the end of 2027. Song said that each generation of Meta's custom chips will take on more technical challenges in exchange for higher performance, including improving performance per unit of power and performance per unit of cost. This also means that, against the backdrop of continuously expanding AI infrastructure investment, Meta is trying to improve computing efficiency and reduce long-term operating costs through self-developed chips. Jointly designed with Broadcom Inc., manufactured by Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR; deployment scale will exceed 1 gigawatt Meta is currently working with Broadcom Inc. (AVGO.US) to develop custom AI chips, with Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR (TSM.US) responsible for production. Song disclosed that Meta has committed to deploying more than 1 gigawatt of related chips within 12 months. If AI demand continues to remain strong, the company expects the pace of deployment thereafter to further accelerate. At present, Arke has already entered actual testing. On September 1, the first batch of 12 Arke chips was delivered to Meta by Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR, with the gap between actual performance and previous simulation results being only 2% to 3%. More notably, this batch of processors has already successfully run Meta's own AI models, and also ran models from DeepSeek and Alibaba Group Holding Limited Sponsored ADR (BABA.US), showing that Meta's self-developed chips are not only capable of being optimized for a single internal model, but have the ability to run different AI models. Giving up the "two-pronged" approach of training and inference; Meta shifts the focus of its self-developed chips to AI inference Meta's self-developed AI chip strategy has also undergone adjustments. The company had previously planned to develop a chip called Olympus, hoping to use it for both AI model training and inference, but later canceled the project, partly due to cost considerations. After that, Meta focused its self-developed chips more on AI inference. Song said: "These will become our main chips for general inference." As Meta continues to integrate AI features into Facebook, Instagram, and other products, model invocation volume continues to expand, and the computing resources required for inference also increase accordingly. Therefore, compared with simply pursuing higher computing power, reducing the cost and energy consumption of each AI inference is becoming increasingly important to Meta. Meta's self-developed chip strategy has therefore gradually formed a clearer direction, namely not trying to immediately cover all AI computing tasks, but first building dedicated chips for large-scale, continuously occurring inference workloads, reducing the overall operating costs of AI infrastructure by improving performance per watt and performance per dollar. Competition in AI infrastructure extends to self-developed chips; Meta says it has a "very robust" product roadmap As large technology companies continue to expand AI infrastructure investment, self-developed chips are becoming an important means of controlling AI costs and improving data center efficiency. Meta's advancement of MTIA 450 and MTIA 500 this time shows that its self-developed chip project is moving further from the testing stage toward large-scale deployment. Under the current plan, MTIA 450, which is being tested, will enter data centers in the first half of 2027, while the next-generation MTIA 500 is expected to be deployed before the end of 2027. Future generations of products will further improve operating speed and throughput. Song said that Meta has already formulated a "very robust roadmap" for custom chips.