Stock Surges Over 3.5%! Google Developing New 'Frozen v2' Chip to Boost AI Model Performance

According to the latest report from foreign media, Google is quietly developing a new server chip internally codenamed 'Frozen v2.' This chip aims to permanently embed the underlying computational architecture of the Gemini large model directly into silicon hardware, significantly addressing the current severe shortage of AI computing power. Following this news, Alphabet, Google's parent company, saw its stock price surge by over 3.5%.

U.S. tech media The Information reported on Monday (the 20th), citing sources, that the chip is expected to be deployed as early as 2028. However, it is not intended to replace Google's existing general-purpose Tensor Processing Units (TPUs), but rather to serve as a complementary new line of specialized products.

The name 'Frozen' stems from its core design philosophy: Unlike NVIDIA GPUs or Google TPUs, which require real-time execution of massive logical operations to adapt to multiple models, Frozen v2 directly embeds the underlying computational logic of the Gemini series, drastically reducing computational steps and data movement.

According to calculations by Google's R&D team, once deployed, Frozen v2 is projected to process 6 to 10 times more tokens per unit of power consumption than Google's current latest in-house AI chip. This represents a qualitative leap in energy efficiency, enabling faster user query responses and supporting new AI applications.

However, specialization comes at the cost of flexibility. Frozen v2 assumes that future Gemini models will maintain the current underlying architecture. A major architectural change could render the chip inoperable.

The report notes that Google's approach resembles that of Canadian startup Taalas. The initial concept was led by DeepMind's Chief Scientist Jeff Dean but was shelved for years due to the hardware's short lifecycle, as it could only support a single version. Frozen v2 seeks balance by adjusting the depth of hardware fixation.

Insiders indicate that the computing power gap has already triggered internal resource conflicts at Google, forcing Google Cloud to reject some external customer orders.

Currently, Google has no plans for large-scale mass production and views Frozen v2 as a technology testbed.

In response to the report, a Google spokesperson stated, 'Our teams continuously develop various innovative technologies to achieve extreme performance. Not every project will reach mass production, but this exploration is at the heart of our full-stack in-house development strategy.'

Currently, companies including Samba Nova, OpenAI, Microsoft, and NVIDIA are all increasing investments in specialized inference chips, intensifying the competition in software-hardware integration for AI infrastructure.