Google is working on new ‘Frozen v2’ AI chip for Gemini: Here’s what it could do
AI-summarised brief · reviewed before publication
Google is developing a new custom AI server processor, codenamed “Frozen v2,” to accelerate inference for its Gemini models. Unlike the existing Tensor Processing Unit line, the chip will embed selected Gemini components directly in silicon, aiming to cut power use and boost the number of requests processed per watt. Engineers are still deciding how much of the model to hard‑wire, and the design is slated for possible deployment as early as 2028. The initiative responds to rising inference costs and cloud capacity constraints that have limited service availability for some Google Cloud customers. If the chip achieves its projected six‑to‑ten‑fold efficiency gain over current hardware, it could expand Google’s ability to serve enterprise AI workloads while reducing reliance on third‑party processors.
💡 Why It Matters
- · Embedding AI logic in hardware could dramatically lower operating expenses, giving Google a cost‑advantage as generative‑AI demand surges.
