The scalability of new AI models is increasingly constrained by current architectures, particularly the limitations imposed by the number of GPUs that can fit inside a single rack. Moving optical attach directly to the GPU enables scale-up domains of more than 512, overcoming these bottlenecks and providing a path for the next generation of foundation models. Over the past five years, Broadcom’s significant investment over the past five years in Co-Packaged Optics (CPO), is paving the way for optical I/O from the GPUs at cost parity to traditional copper interconnects. The presentation will highlight the benefits of Broadcom’s development strategy, focusing on key decisions such as edge coupling and external lasers, core technology innovations such as a detachable fiber connector, and large-scale manufacturing through automated assembly and test.