MOUNTAIN VIEW — Google is actively developing a specialized artificial intelligence microchip designed specifically to power its flagship AI model, Gemini. According to multiple industry sources cited by tech publication The Information on July 20, the custom silicon project—internally codenamed "Frozen v2"—targets commercial server deployment by 2028.

Google expects that deploying "Frozen v2" will boost token processing throughput per unit of power by 6 to 10 times compared to existing hardware while drastically reducing response latency for Gemini operations.

Unlike Google’s Tensor Processing Units (TPUs), which are built as general-purpose accelerators capable of running various AI architectures similarly to Nvidia's Graphics Processing Units (GPUs), "Frozen v2" represents a tailored hardware approach. The new chip directly hardcodes Gemini’s data processing mechanisms, model parameters, and neural weights straight onto the silicon layer.

The innovative hardware concept was reportedly proposed by Jeff Dean, Google’s Chief Scientist. The moniker "Frozen" stems from the engineering design principle of locking or "freezing" the structural parameters optimized for Gemini into fixed hardware pathways. However, Google executives and engineers are still deliberating the precise degree of specialization versus flexibility to ensure the chip balances high performance with adaptability to future model updates.

Industry analysts point to severe AI compute resource shortages as a primary driver behind Google's bespoke hardware initiative. The company recently faced internal conflicts regarding compute allocation across competing projects, which also constrained its capacity to serve enterprise cloud clients.

To mitigate this compute capacity squeeze, Google recently signed a major infrastructure agreement with Elon Musk’s xAI, paying $920 million monthly (approx. 1.36 trillion KRW) to lease external data center capacity. Furthermore, Google reportedly restricted compute access for rival Meta, which relies on Google Cloud infrastructure to run its AI workloads.

Because "Frozen v2" is tailored to specific iterations of the Gemini architecture, Google does not currently plan mass-market volume manufacturing, as significant structural shifts in future Gemini versions could render fixed hardware obsolete.

A Google spokesperson stated, "We continuously research and experiment with novel technology architectures to deliver peak performance and maximum efficiency for our users and enterprise customers."

Following the revelation, shares of Google’s parent company, Alphabet Inc., rallied nearly 4% during intraday trading before settling up approximately 1.8% around $353 per share in New York.

[Copyright (c) Global Economic Times. All Rights Reserved.]