Meta Platforms plans to deploy its proprietary AI accelerators in data centers starting in the first half of 2027. The company aims to reduce energy consumption and operational costs by decreasing reliance on Nvidia hardware. The development program involves collaboration with Broadcom for design and TSMC for manufacturing.
The company is currently testing the third-generation MTIA 450 accelerator, also known as Arke. Design of the next-generation MTIA 500, codenamed Astrid, is expected to conclude in October 2026, with deployment scheduled for late 2027. Meta intends to reach a total power capacity of over one gigawatt for its accelerator fleet within a year.
On September 1, 2026, TSMC delivered 12 new units to Meta. Initial tests showed performance deviations of only 2-3% from simulations. These chips were used to run Meta, DeepSeek, and Alibaba models. The company clarified that these accelerators are designed for general-purpose inference rather than low-latency tasks.
Meta has officially canceled the Olympus project, which aimed to create a universal accelerator for both training and inference. The decision was driven by financial considerations, as the projected cost for the dual-purpose chip was 30% higher than specialized solutions. The company is now focusing on cost-effective, high-efficiency hardware.
