Alphabet (GOOGL, Financials), the parent company of Google and a major provider of search, cloud and artificial intelligence services, rose about 1% in premarket trading after a report that it is developing a new server chip for Gemini.
The chip, known internally as Frozen v2, would be designed to work closely with Googles AI models, according to The Information.
Employees involved in the project believe it could deliver six to 10 times more tokens per unit of power than Googles latest tensor processing units.
That efficiency would matter as AI companies face rising electricity costs and limited data center capacity. Running models now requires as much attention to power use as raw computing speed.
Google has spent years developing its own chips to reduce its reliance on outside suppliers and improve the economics of AI workloads.
The company introduced its eighth-generation TPUs in April. One version was built for training models, while another focused on inference, the process of running trained systems.
Frozen v2 may not be ready before 2028, leaving time for the design and performance claims to change.
Investors will watch for more detail on the chip and whether Google can turn lower power use into stronger cloud margins.