Meta, Broadcom Extend AI Chip Partnership Through 2029, Plan 2-Nanometer MTIA Deployment

Meta is expanding its in-house “Iris” AI chip program and computing infrastructure as it seeks to reduce reliance on a single GPU supplier and improve efficiency across AI workloads. The strategy reflects a broader shift among cloud giants toward using custom ASICs alongside GPUs, potentially creating new design and manufacturing opportunities for Taiwan Semiconductor Manufacturing Co. and Taiwan’s semiconductor IP and ASIC suppliers.

The latest report said Meta is pressing ahead with Iris and related infrastructure as demand for AI computing remains robust and major technology companies show little sign of slowing near-term capital spending. Analysts expect the parallel expansion of GPUs and custom silicon to reshape Taiwan’s semiconductor supply chain. The report did not disclose Meta’s investment amount, an Iris production target or a firm deployment date.

As demand for AI computing surges, technology giants are investing heavily in proprietary chips to reduce their reliance on suppliers such as Nvidia and AMD. Meta is developing a dedicated AI chip to improve the performance of AI recommendation algorithms and generative applications across its platforms while cutting high hardware procurement and operating costs. The initiative could have significant implications for the global chip supply chain and competitive landscape.

Meta's latest internally developed AI chip, code-named Iris, is scheduled to enter mass production in September 2026. The company is working with chip designer Broadcom, while foundry leader TSMC will manufacture the chip. The move is intended to reduce Meta's reliance on external chip suppliers such as Nvidia. Meta also expects to double its computing capacity in 2027 as it accelerates the deployment of proprietary chips to take greater control of its AI infrastructure.

Meta continues to develop its Meta Training and Inference Accelerator (MTIA) to reduce its reliance on Nvidia AI accelerators, while Samsung Electronics is using its 2nm process to pursue orders from major cloud providers. If Anthropic follows suit with a custom chip, the move could reshape the AI chip supply chain and affect the advanced-process competition between TSMC and Samsung.

As of July 20, 2026, market reports indicated that Samsung Electronics would mass-produce Meta’s third-generation MTIA using its 2nm process, with the order estimated at more than 10 trillion won. Anthropic is also reportedly evaluating Samsung’s 2nm technology for a custom chip, but neither company has announced a formal order, production date or shipment schedule.

Meta is continuing to expand its in-house Meta Training and Inference Accelerator (MTIA) chip lineup to reduce its reliance on third-party AI accelerators. The series supports generative AI and content recommendation systems, while giving Meta flexibility to allocate computing capacity between its own hardware ecosystem and external suppliers.

Meta most recently announced that it is working with Broadcom and TSMC to develop four chips: the MTIA 300, 400, 450 and 500. It has also shortened the product iteration cycle to about six months. The four chips are expected to be progressively rolled out and fully deployed by 2027. Meta will continue buying third-party chips as part of a strategy to diversify supply and computing risks.

Meta continues to expand its data centers to support AI services including its Llama models and advertising recommendations. As generative AI shifts from training to inference at scale, access to GPUs and CPUs has become a key competitive factor. Meta expects capital expenditure of $115 billion to $135 billion in 2026, making secured chip capacity critical to the pace of its growth.

Meta announced a multiyear agreement with NVIDIA on February 17, 2026, without disclosing its value. The company plans to deploy millions of Blackwell and next-generation Rubin GPUs while expanding its use of Grace and Vera CPUs. The long-term deal shows that Meta is securing supplies of AI accelerators while also increasing its allocation of general-purpose processors as inference workloads grow.