Definition
Frozen v2 is the reported internal codename for a google server inference-chip that hardwires parts of the gemini neural-network architecture into silicon while keeping model weights updateable.
Key Points
- Reported by The Information (July 2026); amplified by CNBC, TNW, Reuters/Bloomberg Law citations — not confirmed by Google
- Claimed engineer projections: 6–10× (some outlets 6–8×) tokens per watt vs latest TPUs
- Parallel track to TPU 8t/8i — not a TPU replacement; Gemini-family inference focus
- Deployment target as early as 2028; design still finalizing how much architecture is hardwired
- Framed as response to internal compute-scarcity / Cloud capacity crunch
Warning
Unverified leak. Google spokespersons say teams experiment with efficiency ideas and not every lab project reaches production. Treat efficiency numbers and 2028 timeline as attributed claims.
Contradiction
Efficiency ranges differ across outlets (6–10× TNW/Yahoo/CNBC/TechTimes vs 6–8× NDTV Profit) — inconsistency unresolved without primary confirmation.
Related
- model-specific-inference-silicon
- gemini
- inference-chip
- gpu-infrastructure
- compute-scarcity
- custom-ai-silicon
- jalapeno-chip
- sohu-chip