Google has built a custom AI chip called Frozen v2, designed specifically for its Gemini models. The company says the chip delivers a 6-10x efficiency improvement over its current TPUs. Alphabet shares rose 3% following the announcement.
What Frozen v2 does
The chip is purpose-built for Gemini, Google's next-generation AI model family. It targets a massive jump in performance per watt compared to the Tensor Processing Units that power today's cloud AI workloads. That kind of efficiency gain could cut costs and energy use for training and running large models.
Google hasn't said which Gemini models will use Frozen v2 first, or when the chip will be deployed in its data centers. The company also didn't disclose manufacturing partners or process node details.
Why efficiency matters
AI chips are a hotly contested space. Nvidia dominates with its GPUs, but Google, Amazon, and Microsoft all design their own custom silicon. The goal is to squeeze more performance out of every watt — especially as models grow larger and training costs balloon. A 6-10x improvement over already-optimized TPUs would be a significant leap.
Google's TPU line has been a key part of its cloud AI strategy. The Frozen v2 name suggests this is a second-generation custom chip, though the company hasn't detailed the first Frozen chip publicly.
Market reaction
Investors liked what they heard. Alphabet's stock climbed 3% on the day of the announcement, adding billions to the company's market value. The move signals confidence that Google can keep pace in the AI hardware race without relying solely on outside suppliers.
Still, the chip is not yet in production. Google faces the same supply chain and design challenges as every other chipmaker. The company will need to deliver on the efficiency promises before the gains show up in earnings reports.
No release date has been set. Google also hasn't said whether Frozen v2 will be offered to cloud customers or kept for internal use only.
