Alphabet is designing a server chip for Gemini that could launch in 2028 and be 6x to 10x more power-efficient than Google’s current AI chips, measured by tokens generated per unit of power, according to a report that cited anonymous sources.

Google did not directly confirm the report and said only that its teams are researching and experimenting with new projects.

A 2026 chip comparison said Google’s current Tensor Processing Unit for Gemini API traffic, Ironwood, or TPU v7, was internally deployed for large-scale inference.

Tokens per watt link energy use to AI output, making the metric a practical measure of inference cost and infrastructure efficiency.

AI firms including OpenAI and Anthropic are also pursuing custom chips to ease computing shortages and reduce reliance on Nvidia.

The report comes as Alphabet separately guided US$175b to US$185b in 2026 capital spending and cited compute capacity and power constraints as bottlenecks.

Alphabet, Google’s parent company, shares rose about 3% on July 20 following the report.

🔗 Source: TechCrunch