Google's Frozen v2 hardcodes Gemini for 6-10x efficiency

google is developing 'frozen v2', a server chip embedding gemini's architecture directly into silicon to achieve 6-10x greater efficiency than current tpu chips. the design freezes the model's blueprint into hardware to reduce compute steps, allowing new weights to be loaded while cutting down on data movement. scheduled for deployment in 2028 as a specialized internal solution, this move signals a shift toward domain-specific hardware to optimize margins. if successful, it could allow google to run powerful models at lower prices, increasing competitive pressure on openai and anthropic.