Google is baking Gemini AI directly into silicon to stop begging Nvidia for chips
The search giant Google is working on a secretive hardware project called Frozen v2. Instead of building flexible chips, they are physically engraving Gemini into hardware. It is incredibly smart, but what happens when the software gets outdated?
Internal documents reveal that Google is designing a highly specialized server chip under the codename Frozen v2. Unlike traditional, general-purpose chips that can run any neural network you throw at them, this silicon will have the core math and architecture of the Gemini models permanently hardwired into the physical transistors. It is basically the hardware equivalent of getting a tattoo of your current partner’s name—bold, incredibly committed, and extremely difficult to change.
This hardware-level commitment aims to solve Google's massive headache of energy consumption and hardware shortages. The new silicon is projected to achieve a six-to-tenfold increase in energy efficiency compared to Google's current TPUs. With Google Cloud recently forced to reject some lucrative enterprise clients simply because they ran out of computing power, squeezing more juice out of every watt is no longer a luxury but a survival tactic.
However, this extreme specialization comes with a massive architectural trap. Because the chip's physical layout is tailor-made for Gemini, any future updates to the AI model must strictly play by the hardware’s rules. If Google's researchers invent a revolutionary new AI architecture tomorrow, they will either have to keep using the old logic or throw millions of dollars worth of custom silicon straight into the recycling bin.
The tech giant plans to deploy Frozen v2 as a companion to its existing general-purpose TPUs rather than a total replacement, with the first systems expected to go live no earlier than 2028.
Building custom hardware to run one specific software version is the ultimate flex of corporate desperation in the AI arms race. It shows that the physical limits of our power grids are finally forcing tech giants to choose between software flexibility and actual thermodynamic reality. It remains to be seen if engraving today's AI into solid rock will look like a stroke of genius or the most expensive fossil in computer history.
Source: The Information
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.