MetaAnnounceThe company will expand its strategic partnership with Broadcom, and the two companies will jointly develop multiple generations of...Next-generation custom AI chips—MTIA (Meta Training and Inference Accelerator), which includes the industry's first AI computing accelerator using a 2nm process.
This massive infrastructure project, which extends to 2029, will initially deploy computing power exceeding 1GW (gigawatts). Meanwhile, in response to the rapid expansion of the collaboration, Broadcom President and CEO Hock Tan will step down from the Meta board of directors to become a dedicated technical advisor, personally guiding Meta's customized chip development roadmap.
In the next two years, we will launch the fourth generation of MTIA custom chips, which will fully adopt the Broadcom XPU platform.
In the past, Meta has adopted a diversified chip portfolio strategy when advancing the AI experience of its various applications (such as WhatsApp, Instagram, and Threads). Its self-developed MTIA chip is an accelerator specifically designed for large-scale inference and recommendation systems.
According to a joint statement, Meta expects to develop and deploy up to four new generations of MTIA chips over the next two years, and this deep engineering collaboration is built on the Broadcom XPU (Custom Accelerator) platform. Through this platform, the two companies can tightly integrate logic operations, memory, and high-speed I/O during the chip design phase, ensuring not only the ultimate performance of current AI workloads but also laying a highly adaptable architectural blueprint for future generations of MTIA.
Key to overcoming computing power bottlenecks: Advanced packaging and Ethernet architecture
In AI computing clusters composed of tens of thousands of nodes, the computing power of the chip itself is often not the biggest technical bottleneck; rather, it is the "data transmission" (networking) between nodes.
To support super data centers with power consumption exceeding 1GW, Broadcom not only provides chip design and advanced packaging technology, but also presents its most dominant Ethernet solution.
Through Broadcom's high-radius Ethernet switches, optical connectivity products, PCIe switches, and high-speed SerDes technology, Meta will be able to establish a low-latency, high-bandwidth network architecture. This will completely eliminate network congestion issues that Meta's massive AI computing clusters face when scaling up, scaling out, and scaling across nodes.
Meta CEO Mark Zuckerberg emphasized: "With our initial deployment of over 1GW of custom chips, and as we continue to expand to multiple GW in the future, this collaboration will provide greater performance and energy efficiency for anything we are building, thereby bringing 'personal super intelligence' to billions of people around the world."
To avoid conflicts of interest, Chen Fuyang has been reassigned as a strategic advisor to Meta Silicon.
For the past two years, Broadcom President and CEO Hock Tan has played a key technical role on the Meta board of directors. However, considering the enormous scale of procurement and technology transfer involved in this collaboration, and to avoid potential conflicts of interest, Tan will formally step down from his position as a member of the Meta board of directors.
However, Chen Fuyang has not left Meta's strategic core. Instead, he will serve as Meta's exclusive consultant, continuing to contribute his deep knowledge in silicon chips and system architecture to help Meta shape future infrastructure investments and hardware breakthroughs.
