Overview
Hyperscalers and startups are shifting from general GPU fleets toward custom training/inference ASICs and model-specific silicon amid compute-scarcity and power constraints.
Timeline
- 2026-07: Reported google frozen-v2 Gemini-hardwired inference chip (unconfirmed) (google-frozen-v2-gemini-chip)
- 2026-06: openai jalapeno-chip; etched sohu-chip stealth exit
- Ongoing: nvidia Rubin / Vera Rubin deployments in European sovereign deals (microsoft-mistral-partnership-expansion)
Key Players
Analysis
Architecture lock-in vs tokens-per-watt is the central design tension. Unconfirmed projects like Frozen v2 still move markets as capacity-crunch signals.