Google Is Building a Chip Designed Specifically for Its Gemini AI

Key points

• Google is developing a specialised chip codenamed "Frozen v2" designed to run its Gemini AI models more efficiently, according to a report first published by The Information.
• Google engineers project Frozen v2 could serve 6 to 10 times more AI output per unit of electricity than the company's current custom chips.
• Google is targeting 2028 for deployment of the new chip.
• Google agreed to pay SpaceX roughly $1 billion per month to cover a current shortage of computing capacity while longer-term solutions are built.
• Alphabet's share price rose 3% on Monday after the report emerged.

Google is working on a chip that would be permanently tuned to run one AI model and nothing else. The project is codenamed "Frozen v2", and it targets Google's own Gemini, the large language model (a type of AI that reads and generates text) that powers Google's AI assistant and competes with ChatGPT.

The idea is to bake parts of Gemini's design directly into the processor's circuitry. That means the chip skips steps that a general-purpose chip has to perform every time, cutting the electricity and time needed to answer each query. Google's own engineers project it could handle 6 to 10 times as many responses per watt of power compared with the company's current custom processors, called TPUs (tensor processing units), the specialised chips Google already builds for AI workloads.

That is a striking figure. But it comes with a real constraint.

The chip is essentially frozen around today's Gemini architecture. If Google redesigns Gemini significantly before 2028, Frozen v2 becomes a poor fit. Google reportedly treats the project partly as an experiment and has no current plans to manufacture it at the same scale as its standard TPUs.

Why does this matter to ordinary people?

It matters because computing costs shape what AI services cost, how fast they respond, and how many users they can serve at once. Google is under serious internal pressure right now: the company reportedly had to turn away paying cloud customers due to a shortage of AI computing capacity, and last month it agreed to pay SpaceX close to $1 billion per month to lease capacity and meet its commitments to business clients. A chip that delivers ten times the output per watt would ease that pressure considerably, and lower running costs can eventually translate into cheaper, faster services for end users.

Frozen v2 would sit alongside Google's existing TPUs rather than replace them. General-purpose chips remain essential for training new AI models and handling varied workloads. Frozen v2 would handle high-volume Gemini queries specifically.

Alphabet did not respond to a request for comment before publication. The 2028 target gives Google roughly three years to decide whether the architecture trade-off is worth making permanent.