Overview
Story hub for July 2026 reports that google is developing frozen-v2, a Gemini-architecture-hardwired inference-chip. Unconfirmed by Google.
Recent Developments
- 2026-07-20/21: The Information report; Alphabet shares rose as much as ~3.7% intraday (2026-07-21-google-frozen-v2-tnw)
- Efficiency claims: 6–10× or 6–8× tokens/watt depending on outlet
- Parallel to TPU line; internal Gemini inference; ~2028 target
Warning
Google has not confirmed the project. Spokespersons frame teams as experimenting; not every lab project reaches production.
Contradiction
Tokens/watt claims diverge by outlet: 6–10× (TNW, Yahoo/CNBC, TechTimes) vs 6–8× (2026-07-21-google-frozen-v2-ndtv). Treat as attributed engineer projections, not benchmarks.
Related
- frozen-v2
- model-specific-inference-silicon
- gemini
- compute-scarcity
- custom-ai-silicon
- gpu-infrastructure