Gemini Crosses 1 Billion Users, But Chip Engineers See a Different Hurdle

Last week, while I was debugging a power consumption model for a 5G baseband, a colleague sent me a message: Gemini's on-device inference latency test showed it was only 30% slower than the cloud on Snapdragon 8 Gen 4, but power consumption jumped by 2.5 times. I stared at those numbers for a long time—this isn't a simple optimization problem, but a trade-off for the entire mobile computing architecture. Google announced that Gemini is about to become its tenth product with a billion users, which sounds like a victory for product managers, but as an engineer who saw baseband chips go from tape-out to mass production at Unisoc, I know clearly: a billion users aren't the end point, but the starting point for compute engineering.