NewsNTech
Silicon dependency is the central cost problem in hyperscale AI compute: merchant GPU architectures carry capability that a single operator rarely uses, and hyperscalers pay for all of it.
Google's deal with Marvell, which gives the company the right to acquire up to $12.2 billion in Marvell shares, channels that pressure into a direct investment in custom AI silicon. Marvell's stock rose 6% on the news.
Google and its competitors have been developing custom chips to improve efficiency and reduce their reliance on Nvidia.
The logic of a purpose-built design is that it can dedicate transistor budget to the specific operations a hyperscaler actually runs at volume, rather than carrying the breadth a merchant chip requires to serve thousands of different customers.
Keep reading