Google developing Frozen v2 AI chip with Gemini architecture baked into silicon — 6-10x more efficient than TPUs

What happened

Reports from July 20-21, 2026 reveal Google is developing a new AI server chip internally dubbed Frozen v2, which hardwires parts of the Gemini model architecture directly into silicon instead of running on general-purpose AI accelerators.

Context and impact

Frozen v2 could dramatically reduce inference costs for Gemini models and help Google Cloud address internal compute bottlenecks that have limited enterprise customer availability. The project also signals that Gemini's model architecture is stabilizing enough to be etched into hardware.

Details

• Chip hardwires Gemini architecture elements directly into silicon
• Internal estimates: 6-10x more AI tokens per watt vs. current TPUs
• Target deployment: 2028
• Intended to complement, not replace, the full TPU family
• Partially addresses Google Cloud compute shortages for enterprise
• Alphabet stock reacted positively to the report (CNBC, July 20)