Meta plans to start producing its next-generation in-house AI chip, Iris, in September. The chip is designed by Meta itself, with Broadcom involved in development and TSMC handling manufacturing.
Iris belongs to Meta's MTIA family of custom accelerators and is primarily used for AI tasks behind Facebook and Instagram. It will work alongside GPUs from Nvidia and AMD, not fully replace them.
Iris completed testing in just six weeks with no major issues found. Meta also plans to release a new chip roughly every six months through at least the end of 2027. For comparison, most AI chip refreshes happen on a yearly or longer cycle.
Meta is aggressively expanding its data centers too. It aims to deploy 7 gigawatts of compute capacity this year, doubling to 14 GW by 2027. AI infrastructure spending this year could reach as high as $145 billion.
Building custom silicon helps cut costs and reduces Meta's dependence on Nvidia and AMD. But Meta is still buying large volumes of external GPUs and can't fully cover its own chip needs anytime soon.