Google is reportedly developing a highly specialized artificial intelligence chip designed specifically to run its Gemini models, deepening the company’s effort to control every layer of its AI infrastructure as computing demand rises across consumer products and cloud services.
The server chip, informally codenamed “Frozen v2,” would incorporate elements of Gemini’s architecture directly into hardware rather than relying entirely on a general-purpose AI accelerator. Google expects the design could make serving Gemini between six and 10 times more efficient than using its current Tensor Processing Units, according to people familiar with the project.
Frozen v2 remains at an early stage and is not expected to enter deployment before approximately 2028. Google has not publicly confirmed the project, disclosed manufacturing partners or specified which Gemini models would initially be supported.
The processor is reportedly intended as a specialized addition to Google’s existing custom-chip portfolio rather than a direct replacement for its TPUs. Those accelerators continue to support a broader range of AI training and inference workloads for Google and external cloud customers.
Gemini Architecture Moves Closer to the Hardware
Modern AI accelerators are designed to perform the matrix calculations that underpin neural networks efficiently, but they must remain flexible enough to support different models and rapidly changing architectures.
Frozen v2 would sacrifice some of that flexibility by optimizing silicon for Gemini’s specific computational structure. Frequently used model components could effectively be hardwired into the processor, reducing memory movement, power consumption and the number of general instructions required to generate a response.
The approach resembles an application-specific integrated circuit built around a particular family of AI models. It could substantially lower inference costs if Gemini’s underlying architecture remains stable, but it also creates a risk that future model changes could make portions of the chip less useful before the hardware reaches the end of its operating life.
Google must therefore design the processor around features likely to persist across several generations of Gemini. The long development cycle also means Frozen v2 will need to anticipate models and workloads that may not yet exist.
Compute Shortage Drives Deeper Vertical Integration
The project reportedly emerged amid an internal shortage of AI computing capacity. Demand from Gemini, Google Search, advertising products and Google Cloud has placed increasing pressure on the company’s data centres, while limited capacity has reportedly forced its cloud division to decline some external business.
Google already has one of the industry’s most mature custom-silicon programs. Its latest Ironwood TPU is intended for both AI training and inference, while the company has increasingly offered TPU capacity to outside customers, including major AI developers.
A Gemini-specific chip would push that vertical integration further. Google controls the model, software framework, data-centre network and processor design, allowing engineers to optimize the entire system as one platform.
The strategy could reduce Google’s dependence on Nvidia GPUs and lower the cost of operating Gemini at global scale. However, Frozen v2 is reportedly focused on Google’s own model family, making it less versatile than hardware intended for multiple AI developers.
If the projected efficiency gains are achieved, Frozen v2 could give Google a significant cost advantage in the AI race. For now, it remains an experimental project whose commercial value will depend on whether Gemini’s architecture evolves slowly enough for dedicated silicon to remain relevant when deployment begins.
