The Zhitong Finance App learned that Meta Platforms (META.US) plans to deploy a new generation of self-developed artificial intelligence chips to data centers in the first half of 2027, hoping to reduce the energy consumption and costs required to run AI models through customized chips. At the same time, the company has promised a deployment scale of more than 1 gigawatt for related chips, and stated that if AI demand continues to be strong, the speed of subsequent deployment will be further accelerated.
Yee Jiun Song, vice president of Meta Engineering, said in an interview that the company's third-generation self-developed AI processor, MTIA 450, is currently in the testing phase. The chip is codenamed “Arke”; the next-generation MTIA 500 is codenamed “Astrid,” and the design is expected to be completed in about a month, and it is planned to enter the data center by 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 improved performance per watt and per cost performance. This also means that in the context of increasing investment in AI infrastructure, Meta is trying to improve computing power efficiency and reduce long-term operating costs through self-developed chips.
Jointly designed by Broadcom, TSMC's foundry deployment scale will exceed 1 gigawatt
Meta is currently cooperating with Broadcom (AVGO.US) to develop custom AI chips, which are produced by TSM.US. Song revealed that Meta has promised to deploy more than 1 gigawatt of related chips within 12 months. If AI demand continues to be strong, the company expects further acceleration of deployment in the future.
Arke has now entered the actual testing phase. On September 1, the first batch of 12 Arke chips was delivered by TSMC to Meta. The gap between their actual performance and previous simulation results was only 2% to 3%.
What is more noteworthy is that this batch of processors has successfully run Meta's own AI models, while also running DeepSeek and Alibaba (BABA.US) models, showing that Meta's self-developed chips can not only be optimized for a single model within the company, but have the ability to run different AI models.
Abandon training and reasoning and “grab with both hands”, Meta shifts its focus to AI reasoning with self-developed chips
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 AI model training and inference at the same time, but the project was later cancelled due in part to cost considerations. Since then, Meta has focused its self-developed chips more on AI inference. “These will be our workhorse chips for general reasoning,” Song said.
As Meta continues to integrate AI functions into Facebook, Instagram, and other products, the number of model calls continues to expand, and the computational resources required for inference also increase. Therefore, compared to simply pursuing higher computing power, it is increasingly important for Meta to reduce the cost and energy consumption of each AI inference.
As a result, Meta's self-developed chip strategy has gradually taken a clearer direction. Instead of trying to immediately cover all AI computing tasks, it is first creating dedicated chips for large-scale, ongoing inference workloads to reduce the overall operating cost of AI infrastructure by improving performance per watt and per dollar.
AI infrastructure competition extends to self-developed chips Meta says it has a “very robust” product roadmap
As large technology companies continue to expand investment in AI infrastructure, self-developed chips are becoming an important means of controlling AI costs and improving data center efficiency. Meta is now advancing the MTIA 450 and MTIA 500, showing that its self-developed chip project is moving further from the testing phase to large-scale deployment.
According to current plans, the MTIA 450 being tested will enter the data center in the first half of 2027, while the next 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 Meta has developed a “very robust roadmap” for custom chips.