Tech giants are ramping up in-house AI chip development to cut costs and reduce reliance on NVIDIA, and Meta is following suit. According to Reuters, Meta is testing its first in-house AI training chip, while another Commercial Times report suggests the chip will be built on TSMC’s 5nm process, with mass production set for 2026.
As highlighted by Commercial Times, in addition to TSMC’s 5nm node, Meta’s MTIA v2 chip will also leverage TSMC’s CoWoS technology, with design supported by IPs from Broadcom.
As mass production is set for the first half of next year, it would benefit Taiwanese suppliers like TSMC, Quanta, Asia Vital Components and Chenming Electronic Tech, the Commercial Times report adds.
Meta’s Plans for In-House AI Chips
For now, Meta remains one of NVIDIA’s top customers, using GPUs to train models for ads, recommendations, and its Llama series, while also handling AI inference for over 3 billion daily users, according to Reuters.
However, though DeepSeek’s breakthrough raises doubts about future chip demand, soaring costs and NVIDIA reliance are pushing CSP giants, including Google (TPU), Microsoftc (Maia), Amazon (Trainium) and even OpenAI, to seek alternatives, with TSMC reportedly handling a large portion of the chips.
As for Meta, it is also developing in-house chips to cut infrastructure costs as it bets big on AI. The company projects 2025 expenses of $114–119 billion, with up to $65 billion for AI-driven capital spending, as noted by its press release.
In terms of Meta’s blueprint for its in-house AI chips, Reuters notes that the social media giant began using its MTIA chip in 2024 for AI inference, powering recommendation systems on Facebook and Instagram. Notably, the report indicates that it aims to deploy in-house training chips by 2026, starting with recommendations before expanding to generative AI like Meta AI.
Read more
(Photo credit: Meta)