The chip, internally dubbed "Frozen v2," could serve 6 to 10 times more tokens per unit of power than Google's latest TPUs
Getty Images / Bloomberg
The chip, referred to internally as "Frozen v2," cuts processing overhead by baking parts of Gemini's architecture into the silicon itself, limiting how many calculations the hardware must perform and how far data must travel to generate a response, according to Bloomberg. Engineers working on the project estimate the chip could deliver six to ten times the token output per watt compared with Google's latest TPUs.
Google developed the project in part to relieve a severe internal capacity crunch that has created friction within the company and led Google Cloud to decline business from external customers. Google is targeting 2028 for deployment, though engineers are still finalizing the chip's design and how much model information will be hardwired into it, according to Yahoo Finance.
Frozen v2 would represent a specialized branch of Google's chip portfolio rather than a replacement for its general-purpose TPUs. That efficiency comes at a cost, however: because parts of the model are hardwired into the silicon, the chip can only support future Gemini versions if Google keeps its foundational architecture intact. Production volumes are expected to fall well short of TPU levels, and the effort is for now treated internally as an exploratory exercise rather than a full-scale rollout.
Google has faced mounting pressure on multiple fronts, including the departure of senior AI researchers to rivals. Noam Shazeer, who co-led the Gemini models, announced he was joining OpenAI, while Nobel Prize recipient John Jumper left Google DeepMind for Anthropic. Alphabet stock fell as much as 7% following those departures. The next Gemini Pro release has also been delayed, and Chinese AI models now account for 45% of U.S. company token use.
In a statement to CNBC, Google said its teams are "constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers," adding that "not every project moves into production." A Google Cloud spokesperson said the company's approach involves "co-designing our hardware and software from the ground up" to optimize systems for real-world workloads.
Google is not alone in pursuing custom AI silicon. OpenAI recently introduced its first in-house chip, built alongside Broadcom $AVGO Inc., while Anthropic has begun preliminary discussions with Samsung Electronics Co. about developing its own silicon, according to Bloomberg.
Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.
