ChipsJuly 20, 2026via The Decoder

Google's "Frozen v2" chip reportedly bakes Gemini's architecture directly into silicon for efficiency gains

Why it matters

Google is moving beyond general-purpose TPUs to model-specific hardware, a strategic shift that could lock in cost advantages for Gemini inference and raise the bar for competitors' infrastructure spending.

Key signals

  • Chip codenamed 'Frozen v2'
  • Bakes Gemini architecture directly into silicon
  • Reported 6-10x efficiency gain vs current TPUs
  • Scheduled deployment: 2028
  • Goal: reduce AI inference costs and competitive pricing vs OpenAI/Anthropic
  • Source: internal sources (unverified)

The hook

6-10x more efficient. Google's custom silicon play could reshape AI economics by 2028.

Google is developing "Frozen v2," a server chip that bakes the Gemini architecture directly into hardware. According to internal sources, it could be 6 to 10 times more efficient than current TPUs. Scheduled for 2028, the chip would drastically cut Google's AI inference costs and could give the comp

The week's key stories, every Friday.

For practitioners and enthusiasts — free, in your inbox.

Free forever. No spam.