Google’s reported chip initiative reflects the next stage of the AI race, where competitive advantage increasingly depends on designing hardware and software as a unified system capable of delivering more intelligence with dramatically greater efficiency. (Source: Image by RR)

Google Develops New AI Chip Optimized Specifically for Gemini Models

Google is reportedly developing a new generation of AI server chips designed specifically to run its Gemini models more efficiently by embedding portions of the model directly into the hardware architecture. According to reports, the project—internally referred to as “Frozen v2″—aims to dramatically improve inference efficiency while helping Google address growing AI computing capacity constraints that have reportedly limited its ability to meet customer demand through Google Cloud.

Unlike Google’s existing Tensor Processing Units (TPUs), the new chips are reportedly intended to complement rather than replace the company’s current AI hardware portfolio. Engineers, as noted in an article at reuters.com, are exploring ways to hardwire portions of Gemini directly into the silicon, potentially allowing the processors to serve AI tokens far more efficiently than conventional accelerators. Reports suggest the specialized chips could improve power efficiency by six to ten times compared to Google’s latest custom AI processors, though the design remains under active development with deployment targeted for 2028.

The initiative reflects the increasing pressure AI companies face as demand for generative AI continues to outpace available computing infrastructure. Rather than relying solely on larger data centers and more graphics processors, Google appears to be pursuing tighter hardware-software integration to improve performance while reducing power consumption and operational costs. The approach mirrors a broader industry trend in which frontier AI companies are increasingly designing custom silicon optimized for their own models instead of depending entirely on general-purpose hardware.

More broadly, the project illustrates how the AI race is expanding beyond algorithms into semiconductor architecture itself. As models become larger and more computationally intensive, future competitive advantages may depend less on building smarter AI alone and more on engineering highly specialized hardware capable of delivering greater intelligence with dramatically lower energy consumption. The companies that master both software and silicon could define the next generation of AI infrastructure.

read more at reuters.com