Broadcom is the second best chip company in the world after Nvidia.

It designs custom chips (XPUs) for specific tasks like training, inference, ranking, vector search, and recommendation. They are one of the few companies in the world that can actually design and manufacture these chips at hyperscale.

Companies come to Broadcom because the custom silicon team there does something unique. They work side-by-side with cloud providers and AI labs to design chips around the exact needs of their models and infrastructure.

Compute logic, memory layout, interconnect fabric, packaging, and power delivery are all co-designed and optimized as one system. This lets these companies run their specific AI workloads faster, more efficiently, and at lower cost than using generic chips.

That’s why AWS, Google, Meta, Microsoft, and OpenAI all work with Broadcom in different parts of their AI infrastructure co-developing multiple generations of custom AI chips and networking solutions.

And it’s not an either-or. GPUs and XPUs will grow together. GPUs are the general-purpose engines that train and run large models, flexible and essential for deep learning. But as AI scales, many workloads benefit from dedicated accelerators optimized for very specific tasks. That’s where XPUs come in. XPUs handle high-volume tasks like inference, ranking, recommendation, and vector search at lower cost and higher efficiency. They don’t replace GPUs. Instead they’ll work together.

Nvidia and Broadcom are two of the biggest winners of the AI infrastructure buildout and the new industrial revolution.