Meta, Broadcom strengthen ties to co-develop AI chips

Meta and Broadcom have announced an expanded multi-year strategic partnership to co-develop multiple generations of custom silicon.

This agreement focuses on the Meta Training and Inference Accelerator (MTIA), which is the custom hardware used to power artificial intelligence (AI) across the various apps and services owned by Meta.

The collaboration is a major step for both firms as they seek to build the massive computing foundation required for modern AI at a global scale.

Broadcom is providing its XPU platform, which is a technology designed for creating custom AI accelerators. This platform allows the two companies to tightly couple logic, memory, and high-speed data transfer components for current and future hardware iterations.

Furthermore, Broadcom is delivering advanced Ethernet networking solutions to eliminate bottlenecks in Meta’s rapidly expanding compute clusters. This includes high-speed serialiser-deserialiser or SerDes capabilities, which enable fast data transmission between components.

This partnership reinforces Broadcom’s leadership in the AI networking sector and provides a sustained roadmap through 2029. Meanwhile, Broadcom President and CEO Hock Tan is transitioning from his position on the Meta Board of Directors into an advisor role to provide guidance on the custom silicon roadmap.

Diversified silicon roadmap

Meta is pursuing a portfolio approach to hardware, wherein it uses a variety of different chips optimised for specific tasks. While the company buys chips from outside partners, it keeps its own MTIA silicon at the centre of its strategy to improve efficiency and reduce the total cost of ownership.

The history of this programme shows a rapid evolution. In May 2023, Meta announced the development of the first generation MTIA v1. This was followed in April 2024 when the company introduced the next generation of these custom chips.

The pace of development has since accelerated significantly. This March, Meta stated it was developing and deploying four new generations of MTIA chips within just two years. This is much faster than the typical industry cycle of launching a new chip every one or two years. By releasing new versions every six months or less, Meta can quickly adapt to new techniques in AI.

This internal development is supported by several high-profile partnerships with other industry leaders. In February, Meta announced a long-term partnership with NVIDIA to supply technology for AI-optimised data centres. Shortly after, Meta announced an agreement with AMD to power its infrastructure with up to 6GW of AMD Instinct GPUs. Last month, Meta said it was partnering with Arm to develop a new class of CPUs to support large-scale AI deployments.

Scaling for personal superintelligence

These massive investments are directed towards delivering what Meta calls personal superintelligence to billions of users globally, which is essentially providing an advanced AI assistant capable of handling complex tasks and providing highly personalised experiences.

To achieve this, Meta is scaling its infrastructure to an unprecedented degree. The initial commitment with Broadcom exceeds 1GW, and is the first phase of a sustained, multi-gigawatt rollout.

The financial scale of this ambition is equally significant, with Meta’s capital expenditure for the full year of 2025 touching $72.22 billion. For 2026, the company expects this to rise to between $115 billion and $135 billion, driven largely by its AI and infrastructure investments.

Meta’s strategy involves more than just buying chips as it is also securing the energy and networking capacity needed to keep these systems running. The partnership with Broadcom specifically utilises 2nm technology. Smaller transistors generally allow for better performance and higher energy efficiency.

By co-designing hardware that is tightly integrated with its software stack, such as the PyTorch framework, Meta aims to achieve greater efficiency than would be possible using general-purpose chips alone.

Last week, the social media giant introduced Muse Spark, which is the first in a new series of large language models built by Meta Superintelligence Labs.