Alphabet rises after report on more efficient Gemini server chip

Alphabet shares rose about 3% on Monday after The Information reported that Google is working on a new server chip intended to run its Gemini artificial intelligence models with greater power efficiency. CNBC quote data showed Alphabet Class A stock up 3.38% at $358.50 at 10:10 a.m. EDT.

The chip is known internally as “Frozen v2,” The Information reported. It would place parts of Gemini’s model architecture permanently into the chip itself, a design choice aimed at cutting the amount of computation and data transfer needed when the system responds to user prompts.

That approach differs from Google’s tensor processing units, or TPUs, which are custom AI chips used across a broader set of machine-learning workloads. The Information said Frozen would sit as a more specialized line within Google’s chip portfolio, rather than replacing TPUs.

Efficiency gains and trade-offs

Google engineers estimate the Frozen v2 design could process six to ten times more tokens for each unit of power than the company’s latest TPUs, according to The Information. In practical terms, the reported design seeks to save energy by building fixed elements of the Gemini architecture into silicon, reducing work that would otherwise be handled through more flexible computation.

The efficiency gain would come with limits, The Information said. The chip would be useful for future Gemini models only if Google keeps the same underlying architecture. If the model design changes materially, hardware with fixed architectural elements could become less adaptable than general-purpose AI accelerators.

The Information reported that Google is treating Frozen v2 partly as a test and does not currently plan to manufacture it at the same scale as its TPUs. The company is targeting deployment in 2028, according to the publication.

Compute demand remains a constraint

The reported project comes as Google faces heavy demand for AI computing capacity. The Information said the effort is meant to ease a significant internal compute shortage that has created tensions inside the company and reportedly led Google Cloud to turn away some outside business.

CNBC reported last month that Google had agreed to pay SpaceX nearly $1 billion a month to help close the gap and meet enterprise compute commitments. The Information’s account suggests Frozen v2 is one of the longer-term responses to the same capacity pressure, though the reported 2028 target means it would not address near-term constraints.

Alphabet did not immediately respond to CNBC’s request for comment.

Custom silicon has become a central lever for large technology companies building and operating AI systems, because the cost of serving models depends heavily on energy use, chip availability and data-center capacity. The reported Frozen v2 plan shows Google exploring a narrower hardware design for Gemini, while maintaining its broader TPU program for more flexible AI workloads.

This story draws on original reporting from CNBC.