Google Develops Frozen v2 Chip to Solve Gemini Compute Crunch

## Google is developing a specialized AI inference processor, internally codenamed Frozen v2, designed to embed the structural architecture of its Gemini AI models directly into the silicon. By using hardwired routing and operator fusion rather than computing requirements at runtime, the chip aims to deliver six to ten times more AI tokens per unit of power than current Tensor Processing Units (TPUs). Targeted for data center deployment as early as 2028, the hardware is intended to complement existing TPUs and resolve a critical compute shortage that has forced Google Cloud to reject customer deals—including requests from Meta Platforms—and manage a $462 billion backlog.

To bridge this capacity gap, Google is reportedly paying Space Exploration Technologies Corp. nearly $1 billion monthly to rent approximately 110,000 Nvidia GPUs. Alphabet Inc. has responded to these infrastructure needs by increasing its 2026 capital expenditure guidance to between $180 billion and $190 billion. Shares of Alphabet rose over 3% following reports of the Frozen v2 project, which represents a strategic bet on the stability of the transformer-based architecture to justify a loss of flexibility for extreme efficiency.

Parallel to these hardware efforts, Google released three new AI models on July 21 and 22: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. These launches target cost efficiency and cybersecurity, though the flagship Gemini 3.5 Pro remains delayed past its June launch date after failing to meet internal coding performance goals. The company has confirmed that training for Gemini 4 has commenced as it faces intense competition from OpenAI and Anthropic.