By Natasha Abellard
Publication Date: 2026-01-30 20:29:00
If money were not a factor, a Google software engineer says custom chips, like those from Broadcom , would always be his default selection to build cutting-edge artificial intelligence models. “I am a hundred percent using a custom chip of some sort if resources are not a constraint and I need to move fast and get as much training as possible,” said Gabriel Rasskin, who works on the Gemini AI team. “Every second of compute matters,” he stressed in an interview with CNBC. That mindset appears to be growing among hyperscalers as their appetite for custom chips rises, with Broadcom leading the way. Google used tensor processor units (TPUs), co-designed by Broadcom, to successfully train Gemini 3, whose November launch put the Alphabet unit back on the map as the large language model to beat. The performance of the TPUs also put Google in the conversation as an AI chip alternative to the industry-standard Nvidia GPUs, or graphics processing units. Every second of compute matters. Google Gemini software engineer Gabriel Rasskin Custom chips are designed for specific high-volume tasks, something Nvidia’s general-purpose GPUs are not built to do. Nvidia CEO Jensen Huang, however, recently dismissed custom chips as a threat to his business. “What Nvidia does is much more versatile,” he told Jim Cramer in an interview last month . “Nvidia can address markets that are much, much broader, not just chatbots.” Back in November , Nvidia said it was “delighted by Google’s success” in a…
