Google Targets Gemini Efficiency With New Frozen v2 Server Chip

Google is reportedly designing a new custom server chip, codenamed Frozen v2, to enhance the efficiency of its Gemini AI models. The initiative seeks to overcome existing limitations in AI computing capacity that have previously constrained Google Cloud's ability to onboard additional external customers. By integrating specific AI model elements directly into the hardware, Google aims to improve token performance per unit of power by six to 10 times compared to its current custom AI hardware.

The Frozen project represents a move to develop a specialized family of AI chips that complement, rather than replace, Google’s existing Tensor Processing Units (TPUs). While TPUs currently handle a wide range of internal and cloud workloads, the Frozen v2 chips will focus on optimizing specific AI inference tasks. Engineers are currently in the design phase, targeting a potential deployment window around 2028, as the company seeks to build highly integrated, optimized systems from the ground up.

This development occurs amid intense pressure on tech firms to justify massive capital expenditure on AI infrastructure. While the semiconductor industry has seen sharp gains this year driven by AI demand, investors remain cautious about the long-term sustainability of these hardware costs. For India's growing AI ecosystem, Google's investment in hardware efficiency suggests a long-term commitment to lowering inference costs, which could eventually lead to more affordable and scalable AI services for Indian enterprises and developers using Google Cloud.