Meta and Broadcom: Agreement Extended Until 2029 for New 2-Nanometer MTIA Chips

Meta has announced an expansion of its strategic collaboration with Broadcom, extending the existing agreement for the design of proprietary AI accelerators until 2029.

At the heart of the agreement is the evolution of the MTIA (Meta Training and Inference Accelerator) chip family, which will mark a significant technical milestone: according to Broadcom, these will be the first artificial intelligence chips to utilize a 2-nanometer manufacturing process.

In a market dominated by the expensive and increasingly hard-to-find GPUs from NVIDIA and AMD, Meta is following in the footsteps of Google and Amazon by betting on ASICs (Application-Specific Integrated Circuits). Unlike general-purpose GPUs, the MTIA chips are specifically optimized for Meta's internal workloads, such as inference and recommendation systems. The company has recently unveiled a roadmap that includes four new models over two years.

While cloud competitors offer their chips to external customers, Meta's solutions will remain for exclusive use within its own infrastructure (at least for now), which currently has a pipeline of 31 data centers (27 of which are in the United States). In the initial phase, Meta will install MTIA accelerators with a computational power value of over 1 GW (Gigawatt), but the medium-term goal - by 2027 - is to reach a scale of several Gigawatts. It is noteworthy that Meta plans to spend up to $135 billion on AI this year alone. This initiative complements other recent agreements, including the deployment of 6 GW of AMD GPUs, the purchase of millions of NVIDIA chips, and a deal with Arm.

The announcement has also led to a reorganization of governance. Hock Tan, President and CEO of Broadcom, will leave the Board of Directors of Meta (where he has been since 2024) to take on the role of strategic advisor.