Google Designs Frozen v2 Chip to Cut Gemini Running Costs
Alphabet is building a custom server chip called Frozen v2 to run Gemini models more efficiently. The hardware targets steep power consumption and the ballooning costs of operating AI at scale.
Google is designing a custom server chip, code-named Frozen v2, to run its Gemini artificial intelligence models with significantly higher energy efficiency. The project, first reported by The Information, is part of a broader push by the company to optimize its hardware stack from the ground up. Google aims to deploy the chip in its data centers by 2028.
Frozen v2 could be six to ten times more efficient than Google's existing custom processors, measured by the number of tokens an AI model generates per unit of power. Google did not formally confirm the design details, but a spokesperson said the company regularly co-designs its hardware and software to maximize performance.
Google has been building custom silicon for over a decade, pioneering the Tensor Processing Unit long before most of its peers. Frozen v2 continues that tradition but with a sharper focus: inference efficiency rather than raw training power. Inference refers to the daily work of running live models for actual users. These operational workloads consume far more sustained energy over time than the initial phase of training a model, making efficiency gains here especially consequential.
This development fits a clear pattern across the industry. Major technology companies are actively reducing their dependence on third-party chipmakers, particularly Nvidia, which dominates the AI hardware market. That dependence has driven prolonged supply shortages and elevated costs for model developers trying to serve growing user bases.
Google is not the only company taking this route. In June, OpenAI unveiled Jalapeño, its first custom inference processor developed with Broadcom. Anthropic is reportedly in talks with Samsung to manufacture custom chips for its Claude models. For Google, these investments are tied directly to its capital position. The company's AI infrastructure buildout is projected to cost between $180 billion and $190 billion, making efficiency improvements not optional but essential.
Financial markets responded positively to news of the project. Alphabet stock rose roughly 3% following disclosure of the Frozen v2 initiative, ahead of its quarterly earnings release. Investors appear to be rewarding cost-containment progress as much as raw performance gains, a shift in sentiment worth noting as infrastructure expenses continue to climb.
What This Means for Africa
For African developers and startups, the implications are practical and immediate. Most AI applications built on the continent rely on foreign cloud APIs, including Google's Gemini endpoints, because local infrastructure and the cost of training proprietary models remain out of reach. The problem is compounded by currency exposure: API pricing is set in dollars, while developers earn revenue in naira, Kenyan shillings, or Ghanaian cedis. If Frozen v2 delivers on its efficiency targets and those savings flow through to API pricing, the cost of building with Gemini could fall meaningfully. That would expand access to generative AI capabilities for resource-constrained teams across the continent, the developers least able to absorb current pricing but most in need of the tools.
Proprietary silicon is no longer just an engineering decision; it is the primary economic lever for any company trying to survive the AI margin squeeze.
Source: TechCrunch
Written by
StacyAI-assisted news curation. Every story is reviewed by our editors before publication.
