- Google is reportedly developing a chip called 'Frozen v2' that hardwires the Gemini model's architecture into silicon, rather than loading models onto general-purpose chips.
- The chip could be 6 to 10 times more efficient than Google's latest TPUs, measured by tokens per unit of power, and is targeted for deployment as early as 2028.
- The project is partly a response to an AI capacity crunch inside Google, which has turned away some cloud customers and caused internal tensions.
- Efficiency is the key advantage: the fixed design reduces power consumption and latency, suiting real-time applications like voice assistants.
- The approach sacrifices flexibility—the chip's architecture is locked to Gemini's current design, though weights can still be updated, risking obsolescence by 2028.
- Google is not alone; startup Taalas offers a similar 'Hardcore' chip that prints a model directly onto silicon, claiming high throughput and lower memory costs.
- The chip would be a separate line from Google's TPUs, deepening its self-reliance from Nvidia and spreading chip orders across suppliers.
- Google has not confirmed the project, calling it experimental; it serves as a signal of the industry trend toward fusing specific models with hardware.