Meta's MTIA project has reached its fourth generation, with previous generations primarily used for recommendation systems and internal inference. The Iris chip continues the long-term strategy of reducing external dependency through self-developed hardware, as Zuckerberg's team increases hardware investment to match the AI demands driven by massive user data.

In terms of capital strategy, Meta is mobilizing resources through collaboration with Broadcom and manufacturing by TSMC, significantly increasing capital expenditures to $145 billion for the deployment of 7-14 gigawatts of infrastructure in 2026-2027. The motivation is to gain pricing power in AI computing by replacing some Nvidia/AMD GPUs with self-developed chips, optimizing the cost structure of AI functionalities on Facebook and Instagram.

Similar to the self-developed chip paths of tech giants like Google TPU and Amazon Trainium, Meta is currently in a critical expansion phase transitioning from reliance on external GPUs to vertically integrating AI infrastructure. The self-developed chips, combined with substantial data center investments, create a closed-loop advantage.

Essentially, this represents capital concentration: self-developed AI chips are reshaping the control of infrastructure among tech giants, with the mechanism being massive capital expenditures and supply chain collaborations that lower unit computing costs, shifting industry pricing power from GPU suppliers to platforms with vast users and data, accelerating the oligopolization of AI infrastructure.

ABAB News · Cognitive Law

Self-developed chips + massive capital expenditures are the structural path to controlling AI computing pricing power.
The era of reliance on external GPUs will end; those who integrate vertically first will gain cost advantages.
User scale will ultimately translate into hardware control, rather than perpetual reliance on suppliers.