Google Develops In-House Chip to Power Gemini More Efficiently
Alphabet's reported "Frozen v2" server chip could run Gemini AI models up to ten times more efficiently, deepening Google's silicon ambitions
Synopsis
Alphabet is reportedly working on a new proprietary server chip, codename "Frozen v2," that will be much more efficient in running Alphabet's Gemini AI models. The chip would deliver six to ten times the amount of AI output for every unit of power, adding to Google's long-term custom-silicon efforts as the industry races to consume less power from Nvidia.
AdvertisementWhat's Being Built
According to The Information, Alphabet is working on a new server chip code-named "Frozen v2" that will enable its in-house AI models to be more efficient. The chip will be available sometime around 2028 and may be as much as 6 to 10 times more efficient than Google's current AI chips, in terms of number of tokens produced per unit of power.
The naming has a technical heritage to it; the "Frozen v2" name comes from a concept of Google DeepMind chief scientist Jeff Dean that he wanted to have the model weights being embedded directly in the chip.
AdvertisementWhy Google is building its own silicon
To help companies optimize hardware to their software, custom silicon is chips that are specifically designed to perform a workload instead of general-purpose computing.
Companies based on AI have been trying to make chips of their own to run their own models more efficiently, to get away from Nvidia's monopoly of the AI hardware industry as well as to overcome the global shortage of computing power for the rapidly growing number of AI models.
It seems that demand pressures are a motivating factor. Google's Gemini capabilities are already proving to be just as limited as demand is from partners like Meta.
Google's Measured Response
The company has neither confirmed nor denied the details. Google told TechCrunch that its teams are continuously working on new innovations to provide performance and efficiency, and that it co-designs hardware and software from the ground up ensures it remains optimised and integrated, with real-world workloads.
AdvertisementThe Competitive Landscape
Rivals are running side by side. In June, OpenAI introduced the first custom inference processor, called the Jalapeño. As per reports, Samsung has been in discussions with Anthropic regarding a chipmaking partnership.
Analysts believe that Google's situation is structurally stronger, a difference that grows with each of its own purpose-built AI silicon chips, as Azure remains almost entirely based on Nvidia's GPUs while Trainium and Inferentia have been bought by only a handful of companies.
The market and the financial context
Prices seem to have priced in the efficiencies. The chip announcement coincides with Alphabet's expected investment in its AI infrastructure, which is estimated to range from $180 billion to $190 billion, and marks a gain of around 3% in Alphabet shares after the news.
Why It Matters
A six to ten fold increase could have a significant effect on Google's cost of serving Gemini at scale, which could enable it to beat out other competitors on API pricing as data centres around the world fight power limits as the main constraint on AI growth.
