Google is developing a purpose-built server chip called Frozen v2 that embeds parts of its Gemini AI model architecture directly into silicon. The company expects the design to substantially reduce the compute needed for inference, with sources telling The Information that the chip could be six to ten times more efficient at serving AI responses than Google’s current TPU hardware. Frozen v2 is being positioned as an internal tool to ease capacity pressure and improve margins rather than a broadly marketed product.
What Frozen v2 is designed to do
Frozen v2 departs from the more general-purpose approach of Google’s Tensor Processing Units (TPUs). Rather than acting as a flexible accelerator for many models, Frozen v2 will have aspects of Gemini’s model structure built into the hardware. The name reflects an analogy to “freezing” in machine learning—locking parts of a model so they no longer change—but in this case parts of the model’s blueprint are encoded in silicon.
Embedding model architecture into hardware reduces the number of software-level compute steps required during inference. According to sources cited by The Information, that reduction in processing can translate to a sizeable improvement in efficiency when serving model responses. Google plans to begin deploying Frozen v2 in 2028, per the report.
Design trade-offs: architecture vs. weights
The idea of baking a model into chip silicon has been explored before inside Google. Jeff Dean, chief scientist at Google DeepMind, originally championed an approach that would have embedded model weights—the tuned numerical parameters that determine model behavior—directly into hardware. That design was abandoned because fixed weights would lock the chip to a single model version and risk rapid obsolescence as models evolve.
Frozen v2 takes a more flexible route by hardcoding the model architecture rather than specific weights. In other words, the chip would implement the model’s structural blueprint in hardware while still allowing new weights to be loaded as the model is updated. How much of the architecture will be permanently encoded and what remains programmable has not been finalized, according to the same reporting.
Target use and production scope
Frozen v2 is intended primarily as an internal solution to help Google scale inference capacity more efficiently. Unlike the TPU line, which Google leases to external customers—including Meta—and sells through cloud services and the TPU@Premises program, Frozen v2 is expected to be produced in smaller volumes and restricted to Google’s own infrastructure.
The Information’s sources say Google sees Frozen v2 as a test case for more specialized chips rather than a new external product category. That narrower focus reflects the chip’s dependence on a particular model architecture; it would only be useful so long as Google continues to use the same underlying design.
Business and competitive implications
If the efficiency gains materialize, Frozen v2 could have meaningful effects on Google’s AI economics. Inference costs are a major driver of operating margins for companies running large language models and other generative AI systems. Lower per-query compute costs would let Google operate larger models more cheaply and potentially pass savings to customers or undercut competitors on price.
Google already positions its TPUs as an alternative to Nvidia and has an internal target to capture a portion of Nvidia’s revenue through TPU sales and cloud services, according to The Information. While Frozen v2 is unlikely to be marketed externally in the same way, its internal cost advantages could still bolster Google’s competitive position in model deployment and pricing versus rivals such as OpenAI and Anthropic.
Limits and open questions
The Frozen v2 concept carries clear trade-offs. Hardcoding architectural elements into silicon reduces flexibility in exchange for efficiency. That makes the design sensitive to future changes in model architecture and training approaches. The Information’s reporting notes that the exact balance between fixed and programmable components is still under discussion.
Another practical limit is timeline and scale. Google’s plan to start deploying Frozen v2 in 2028 suggests a multi-year development cycle and indicates the company is treating the chip as a longer-term infrastructure investment. The smaller production volume relative to the TPU family also signals that Frozen v2 will be a targeted internal optimization rather than a mass-market server accelerator.
What to watch next
Key items to monitor include how much of Gemini’s architecture Google chooses to hardcode, the realized efficiency gains in real deployments, and whether the design leads to additional specialized chips for other internal models. Observers should also watch how Google balances the desire for lower inference costs against the risk of being tied to a particular model architecture as models continue to evolve rapidly.
For now, Frozen v2 represents a notable example of a major cloud provider exploring deeper co-design between AI models and hardware. If the approach succeeds, it could shift how large-scale inference is implemented inside hyperscalers, even if the hardware itself remains an internal asset.
Source: The Decoder