Google plans nearly two million new AI chips as it turns to Marvell for custom designs

Editor's take

Google is reportedly commissioning the production of nearly two million custom AI accelerators, a significant ramp-up in its internal chip development. This move signals Google's intent to further diversify its AI hardware supply chain beyond its own Tensor Processing Units (TPUs) and standard merchant silicon, and highlights the escalating demand for specialized AI compute.

This strategic shift matters because it underscores the intense competition for AI infrastructure among hyperscalers. By enlisting Marvell for custom designs, Google aims to optimize for specific workloads and potentially gain a cost or performance advantage, reducing reliance on NVIDIA's dominant GPUs. This also indicates a maturing AI hardware market where custom silicon is becoming a critical differentiator.

Future developments to monitor include the specific technical capabilities of these Marvell-designed chips and their impact on Google's AI training and inference costs. It will also be telling whether this collaboration leads to broader adoption of Marvell's custom ASIC capabilities by other major cloud providers, potentially fragmenting the AI chip landscape further.