Broadcom Inc. and Meta Platforms Inc. officially announced an extension of their strategic partnership on April 22, 2026, to develop the industry’s first 2-nanometer (2nm) artificial intelligence compute accelerator. This collaboration marks a significant technical milestone in custom silicon production, as Meta transitions its infrastructure to support increasingly complex generative AI models and large-scale recommendation systems. The new 2nm AI chip is designed to provide a substantial increase in transistor density and energy efficiency compared to current 3nm and 5nm solutions.
The partnership leverages Broadcom’s established custom application-specific integrated circuit (ASIC) platform and its intellectual property portfolio in high-speed interconnects and memory interfaces. According to technical specifications released by Broadcom, the 2nm accelerator will utilize advanced gate-all-around (GAA) transistor architecture. This design is expected to deliver a 15% performance boost at the same power level or a 30% reduction in power consumption at the same speed when compared to the previous 3nm generation. The chip will also integrate next-generation High Bandwidth Memory (HBM) and Broadcom’s proprietary PCIe Gen7 and 1.6T Ethernet interconnect technologies to facilitate rapid data transfer across Meta’s massive GPU clusters.
Meta’s decision to co-develop 2nm silicon with Broadcom follows the successful deployment of earlier iterations of the Meta Training and Inference Accelerator (MTIA). By moving to the 2nm node, Meta aims to optimize its hardware-software stack for its latest Llama series models. The custom silicon is specifically architected to handle the specific compute requirements of Meta’s ranking and recommendation algorithms, which drive engagement across its social media platforms, as well as the intensive training needs of its multimodal AI initiatives.
Broadcom’s President of the Semiconductor Solutions Group, Charlie Kawwas, stated that the 2nm project represents the pinnacle of the company’s co-design capabilities. He noted that the integration of Broadcom’s high-speed SerDes and optical interconnects within the 2nm package is essential for overcoming the memory wall and power wall currently facing data center operators. Meta’s Vice President of Infrastructure, Alexis Black Bjorlin, confirmed that the first silicon samples are expected to undergo testing by late 2026, with full-scale deployment in Meta’s global data center network slated for 2027.
This announcement solidifies Broadcom’s position as a primary provider of custom AI silicon for hyperscale cloud providers. Meta remains one of Broadcom’s largest customers in the ASIC space, alongside other major technology firms. The 2nm project will be manufactured using advanced lithography processes, with Broadcom managing the physical design, tooling, and supply chain logistics, while Meta provides the architectural requirements and software integration layers.