Google is reportedly developing a new server chip designed to make its Gemini artificial intelligence models significantly more efficient. Internally known as “Frozen v2,” the chip could reportedly deliver between six and 10 times more tokens per unit of power than Google’s existing AI hardware if it reaches production as planned in 2028.
 |
| Credit: Google |
The reported Google AI chip matters because AI performance is increasingly being measured not only by how powerful a model is, but also by how much computing power and electricity it consumes. For Google, a more efficient chip could help reduce the cost of operating Gemini at enormous scale while strengthening the company’s ability to control more of the technology stack behind its AI ambitions.
Google Is Reportedly Designing a New Gemini Chip
According to the report, Alphabet is working on a new server processor internally called “Frozen v2.” The chip is reportedly being designed to improve the efficiency of Google’s in-house Gemini AI models.
The reported performance target is particularly notable. Frozen v2 could generate between six and 10 times more tokens per unit of power than Google’s current AI chips, based on the information reported by The Information.
That figure should be treated as a reported target rather than a confirmed commercial specification. Google has not publicly announced the chip, its final architecture, release plans, or independently verified performance figures.
In a statement to TechCrunch, Google emphasized its broader strategy of developing hardware and software together. The company said its teams are constantly researching new innovations and that co-designing hardware and software allows its systems to be optimized for real-world workloads.
That response is carefully worded. It confirms that Google is actively exploring new hardware projects, but it does not confirm that Frozen v2 will launch in 2028 or that the reported efficiency figures are final.
Why Google AI Chip Efficiency Matters More Than Raw Performance
The most important detail in this story is not simply that Google may be developing another AI chip. Google has been designing its own AI accelerators for years.
The more consequential development is the reported focus on tokens generated per unit of power.
For generative AI services such as Gemini, the cost of inference can become enormous. Every user request requires computing resources, and popular AI systems may process huge volumes of requests simultaneously. A chip that can produce more output while consuming less electricity could therefore affect the economics of running the service.
This is where efficiency becomes a strategic advantage.
Google is investing heavily in AI infrastructure. The company has said it plans to spend between $180 billion and $190 billion, according to the source material. The more money a company commits to AI data centers, chips, and related infrastructure, the greater the pressure to show that those investments can produce sustainable returns.
A more efficient chip would not automatically solve Google's AI spending problem. However, if the reported performance improvements are achieved in real-world workloads, they could allow the company to run more AI services with the same energy and infrastructure resources.
That could be especially valuable as AI companies increasingly confront the cost of operating large models rather than merely training them.
Google’s Full-Stack AI Strategy Is Becoming More Important
Google has an advantage that many AI companies do not: it controls large parts of the technology stack involved in delivering its AI services.
The company develops AI models such as Gemini, designs custom AI hardware, operates data centers, and controls the software systems that connect those components. That gives Google more opportunities to optimize the entire system rather than relying exclusively on third-party processors.
The reported Frozen v2 project fits into that strategy.
Google's statement about co-designing hardware and software is more than a technical explanation. It points to the potential value of custom AI hardware: the chip does not need to be designed as a general-purpose product for every possible customer. It can instead be optimized around the workloads Google expects its own models to perform.
That approach could make custom chips more efficient than hardware designed to serve a much wider range of AI systems.
However, there is also a trade-off. Custom hardware can be highly optimized for specific workloads, but the technology must keep pace with rapidly changing AI models. A chip designed around one generation of model architecture may become less useful if the company's software strategy changes significantly.
The Reported 2028 Timeline Is Still a Long Way Off
The reported release timeline of 2028 is important because AI hardware development is difficult and subject to change.
A project can be delayed, redesigned, or canceled before reaching commercial deployment. Google itself noted that not every project under exploration moves into production.
That qualification matters when considering Frozen v2.
The reported efficiency figure may represent an internal target, an early design estimate, or performance under specific conditions. It should not be interpreted as proof that a finished 2028 chip will deliver six to 10 times the efficiency of Google's current hardware across all Gemini workloads.
The practical performance of an AI processor depends on more than the chip itself. Memory systems, networking, cooling, software optimization, model architecture, and data-center design can all affect the final result.
Still, even the existence of such a project would demonstrate how much attention AI companies are placing on efficiency.
The AI Chip Race Is Moving From Access to Economics
The most important implication of Google's reported chip project is that the AI hardware race may be entering a less glamorous but more commercially important phase.
Early AI competition was heavily focused on access to computing power. Companies needed as many advanced processors as they could obtain to train increasingly capable models.
Now, the challenge is increasingly about the economics of running those models at scale.
A company that can deliver similar AI capabilities using substantially less energy and computing capacity may have a meaningful advantage over a competitor that relies on more expensive infrastructure. That does not mean efficiency alone will determine which AI company wins. Model quality, user adoption, software ecosystems, and capital will continue to matter.
But the reported Frozen v2 target suggests that Google understands a crucial constraint: AI services cannot be judged only by what they can do. They also have to be affordable to operate.
In that sense, the future of AI competition may be shaped as much by the cost of producing each response as by the number of parameters inside the model generating it.
What the Google AI Chip Could Mean for Users and Developers
For users, a more efficient AI chip could eventually support lower operating costs and greater capacity. That could help Google handle more Gemini usage without increasing infrastructure requirements at the same rate.
However, there is no confirmed evidence that a future Frozen v2 chip would directly lead to lower prices for consumers. Companies can use efficiency gains to reduce costs, expand capacity, improve performance, or increase investment elsewhere.
Developers could benefit indirectly if more efficient infrastructure allows Google to offer AI services at a more sustainable cost. The effect would depend on how Google ultimately integrates the hardware into its cloud and AI products.
For the wider industry, Google's approach also reinforces the growing importance of custom silicon. AI companies are increasingly looking for ways to reduce their dependence on a small number of dominant chip suppliers.
That does not mean third-party AI hardware is becoming irrelevant. Custom chips are expensive and technically demanding to develop. But the potential rewards are becoming large enough for major AI companies to pursue their own alternatives.
Google Is Not the Only AI Company Designing Custom Hardware
The reported Google project comes as other major AI companies explore custom silicon.
OpenAI announced its first custom inference processor, called Jalapeño, in June, according to the source material. Anthropic has also reportedly been discussing a chipmaking partnership with Samsung.
These efforts are not identical, and the companies may have different technical and commercial goals. Still, they reflect the same underlying pressure: AI providers want greater control over the hardware required to operate their models.
Google's position is somewhat different because it already has substantial experience designing AI accelerators. The reported Frozen v2 project would therefore represent an evolution of an existing strategy rather than a completely new direction.
The key question is whether future generations of custom hardware can deliver enough efficiency gains to justify the enormous investment required to design and deploy them.
Google is reportedly developing a new AI chip called Frozen v2 that could make Gemini workloads between six and 10 times more efficient in terms of tokens generated per unit of power. The project and its reported specifications remain unconfirmed, and Google's response indicates that development projects can change before reaching production.
If the reported 2028 target becomes reality, the chip could help Google address one of the most difficult problems in generative AI: the cost of operating powerful models at massive scale.
The broader lesson is more specific than simply saying that Google is building its own hardware. AI competition is increasingly becoming a battle over efficiency. The companies that can turn expensive computing power into more useful AI output may gain an advantage that is less visible than a new model launch, but potentially more important to the long-term economics of the industry.
