NVIDIA is reportedly planning to establish a new business unit that will focus on designing semi-custom chips for its largest data-center customers, according to Reuters. While NVIDIA currently dominates the AI HPC processor market, some of its biggest customers are looking to develop their own AI processor designs and are considering options outside of NVIDIA's offerings. To address this, NVIDIA is creating the new unit.
A semi-custom chip is not simply a chip designed to meet a customer's specifications. It is a chip that is co-developed by NVIDIA and the customer, utilizing NVIDIA's IP blocks and potentially incorporating third-party IP blocks desired by the customer. Additionally, the customer may work with semiconductor fabrication companies like TSMC, Samsung, or Intel Foundry Services for wafer allocation, separate from NVIDIA. For instance, a company like Google may have a pre-allocated amount of wafers with TSMC for its Tensor SoCs used in Pixel smartphones, which it may want to utilize for a semi-custom AI HPC processor for its cloud business. NVIDIA estimates a total addressable market of $30 billion for this specific business unit, encompassing all of its current customers who wish to pursue their own AI processor projects and will now be incentivized to remain with NVIDIA.