Google's 'Frozen v2' Chip Bakes Gemini Into Silicon for 6-10x Efficiency

On July 21, 2026, Alphabet confirmed it is designing a new server chip called "Frozen v2" that is custom-built for its Gemini AI model. The chip bakes part of Gemini's underlying architecture directly into the silicon, so the hardware can skip many decision steps a general-purpose chip would normally make at run time. Per the report, the design is six to ten times more efficient than Google's latest TPU when measured in tokens generated per watt of power. The move is driven by a sharp shortage of AI compute inside Google. Internal tensions over limited resources have grown so serious that Google Cloud has already turned away some outside customers. A dedicated inference chip would let Google run Gemini more cheaply and reduce its dependence on Nvidia, whose GPUs still dominate the AI chip market. Investors reacted quickly: Alphabet's share price rose about three percent the day the story broke. Frozen v2 sits next to, not in place of, Google's TPU line. The existing TPUs, now in their eighth generation, stay general-purpose and can run many different models. The new chip is more specialized: it only works with future Gemini versions that share the same core architecture, though model weights can still be updated. It grew out of an earlier "Frozen" project led by DeepMind's Jeff Dean, which was shelved as too short-lived. Google is not alone. Canadian startup Taalas has raised more than 200 million dollars to bake models into chips, while OpenAI released its own inference processor, "Jalapeño," in June. Nvidia responded by striking a 20-billion-dollar deal with inference-chip startup Groq last December. If Frozen v2 ships as planned in 2028, it will mark a clear shift: the world's biggest AI labs no longer want generic GPUs — they want silicon that knows their model by heart.