Alphabet rises after report of a more efficient Gemini AI chip

Alphabet shares climbed Monday after The Information reported that Google is developing a new server chip aimed at running its Gemini artificial intelligence models more efficiently. For everyday investors, the report points to one of the biggest cost pressures in AI: the electricity and computing power needed to answer user and business queries.

Alphabet’s Class A stock was recently up 3.38% at $358.50 at 10:10 a.m. EDT, according to CNBC market data. CNBC said the stock had risen about 3% following the report.

The chip is internally called “Frozen v2,” according to The Information. The outlet reported that the design would place parts of Gemini’s architecture directly into the silicon, which could reduce the amount of calculation and data movement required when the model responds to prompts.

In plain terms, Google is trying to make the hardware more closely match the software. A general-purpose AI chip can handle many types of workloads. A more specialized chip can be faster or use less power for a narrower task, if the model it supports stays compatible with the chip’s design.

What the chip could change

The Information reported that Google engineers estimate Frozen v2 could process between six and ten times more tokens per unit of power than Google’s newest tensor processing units, or TPUs. Tokens are the small pieces of text, code, or data that AI models read and generate. TPUs are Google’s custom AI chips built to speed up machine-learning workloads.

That metric matters because AI companies pay for both compute capacity and electricity. If a chip can serve more tokens using the same amount of power, it could help lower the cost of running AI products at scale, assuming the technology performs as expected and reaches production.

The Information said Frozen would be a more specialized part of Google’s custom-chip lineup rather than a replacement for TPUs. Google is targeting deployment in 2028, according to the report.

Why Google is pursuing specialized compute

The project is intended to help relieve a major internal compute shortage, The Information reported. The outlet said that shortage has contributed to tensions inside Google and has reportedly pushed Google Cloud to turn away some outside business.

CNBC reported last month that Google agreed to pay SpaceX nearly $1 billion a month to help close the gap and meet enterprise compute commitments. Enterprise compute refers to processing capacity sold to businesses, including cloud customers that need chips and servers to run demanding workloads.

The reported Frozen v2 plan also comes with a constraint. The Information said the chip would support future Gemini models only if Google keeps the same underlying architecture. If Gemini changes too much, a chip with parts of the current architecture embedded into the silicon could lose some of its usefulness.

Because of that trade-off, Google currently sees Frozen v2 partly as a trial run and does not plan to produce it at the same scale as its TPUs, according to The Information.

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

This story draws on original reporting from CNBC.