Meta to deploy custom Arke and Astrid AI chips in 2027: Everything to know

Meta is reportedly planning to deploy its MTIA 450 custom AI chips in its data centers-a significant move to reduce the expenditure required to run AI models. Meta is testing the third-generation of a chip which is called Arke.

The primary motive behind the recent move is to support Meta’s goal of achieving 14 gigawatts of computing capacity by 2027 while still purchasing billions of dollars in Nvidia and AMD GPUs to diversify in hardware suppliers.

Meta is testing its third-generation chip, Arke, specifically designed for training tasks while work on the next processor, MTIA 500 or Astrid is on track to finish next month ahead of its broader deployment by the end of 2027. The deployment adds another compute node as Meta commits tens of billions of dollars to data centers and supporting infrastructure.

Unlike custom silicon, it is purpose-built for sale through a cloud platform, Meta MTIA is designed for internal workloads. The company said that the MTIA 450 delivers six times as many MX4 FLOPS compared to FP16 leveraging its computational speed to increase throughput while preserving model quality. 

MTIA generations will also increase Meta’s high-bandwidth memory (HBM) supply. Furthermore, investors responded positively with gains in Meta call options as they remained optimistic about efficiency boosts.