Google's Secret Weapon: A Custom Chip to Make Gemini Cheaper to Run

Reporting from TechCrunch, Alphabet is reportedly building a new AI chip designed specifically to run Gemini models more efficiently — with the goal of reducing what it costs to serve billions of inference queries every day. The project targets inference, not training: the ongoing, per-query expense that makes running a large-scale AI system an enormous and continuous financial drain.

Google holds an advantage here that most competitors lack: a vertically integrated hardware-to-software stack that lets it design silicon tuned specifically for its own models. A chip purpose-built for Gemini inference could reduce per-query costs considerably, giving Google room to price AI services more aggressively while rivals continue paying standard rates for compute.

The stakes extend beyond cost efficiency. If Google can demonstrably undercut competitors on cost-per-query, it gains a structural advantage in the race to make AI services profitable — a race the entire industry is still running. Smaller players that can't build or afford custom silicon face a harder road if Google uses this chip to move prices.

For Alphabet, the chip represents one more step in the long effort to turn AI from a research expense into a business.

Read the full story at TechCrunch AI