Google developing custom AI chip to boost Gemini efficiency by up to tenfold

Alphabet is developing a custom server chip intended to run its Gemini AI models with significantly higher efficiency, marking the company's latest move to reduce reliance on third-party semiconductor suppliers.

The Frozen v2 project#

The project, called "Frozen v2" internally, is scheduled to arrive in 2028 and could deliver between six and ten times greater efficiency than Google's current AI processors—measured by the volume of tokens the system produces per unit of power consumed. While Alphabet declined to directly confirm the development effort to TechCrunch, the company acknowledged its ongoing hardware innovation work.

Why custom chips matter now#

The race to build proprietary silicon reflects a broader industry shift. As concerns mount about the rising costs of AI infrastructure, technology firms have pivoted toward vertical integration—controlling both the software and hardware layers. This approach helps companies optimize performance while reducing their dependence on Nvidia, which has long held sway over AI chip availability. OpenAI unveiled its Jalapeño inference chip in June, and Anthropic is reportedly negotiating a chipmaking partnership with Samsung, underscoring how central custom silicon has become to competitive AI development.

The financial pressure#

Alphabet faces mounting pressure to demonstrate returns on massive AI investments. The company plans to spend between $180 billion and $190 billion on its AI buildout, and investors have grown increasingly scrutinous about whether such expenditures will translate into profitable returns. Word of the more efficient Frozen v2 design appears to have resonated with markets—Alphabet's stock climbed approximately 3% following the initial reporting, suggesting investor optimism ahead of the company's upcoming earnings announcement.