Google Frozen v2 AI chip aims to cut NVIDIA reliance in 2028

Google is reportedly building the Frozen v2 AI chip to make Gemini models run with 6-10x more tokens per watt than its current hardware. The effort matters because AI compute shortages are still tight, and Google is trying to reduce how much it depends on NVIDIA GPUs for future workloads.

Google Frozen v2 accelerator targets faster Gemini inference and voice workloads

The Frozen v2 AI chip is said to be aimed at real-time AI voice assistants as much as general model work, which gives it a narrower role than a broad training chip. That focus matters because inference is the day-to-day job that powers user-facing AI features, and even small efficiency gains can add up quickly at Google scale.

If the reported efficiency numbers hold, the Google Frozen v2 accelerator could help Gemini handle more daily requests without putting the same pressure on power use and infrastructure. In simple terms, more tokens per watt would mean the system can do more useful work for each unit of energy, which is exactly the kind of improvement large AI operators are chasing right now.

The report suggests Google is not only looking at raw speed, but also at how the chip fits into a live production stack. That includes voice workloads, where latency and consistency tend to matter more than headline benchmark numbers. For a company like Google, a chip tuned around those needs can be more useful than a general-purpose accelerator built for broader training tasks.

It is still important to keep the uncertainty in view. Google has not confirmed the report, and the details around performance and deployment are not official. Even so, the claim points to a clear direction: custom silicon that is built around Google’s own models, rather than a one-size-fits-all hardware strategy.

Alphabet 2028 chip launch comes as AI hardware spending comes under pressure

Google’s reported chip effort also lands at a time when AI hardware spending is under close watch. Alphabet is already guiding toward $180 billion to $190 billion in capex, so any move that improves efficiency would be seen through the lens of cost control as much as technical progress.

That is one reason the Frozen v2 AI chip report has drawn attention. If Google can improve inference efficiency while relying less on external NVIDIA hardware, the company could ease some of the pressure that comes with large-scale AI infrastructure spending. This does not mean NVIDIA dependence disappears, but it does suggest Google wants more control over the economics of running Gemini.

Seen that way, a successful Alphabet 2028 chip launch would not just be about launching another accelerator. It would be a sign that Google’s full-stack hardware approach is still central to its AI strategy. In other words, the company appears to be betting that custom silicon can support its models more efficiently than renting that performance from outside suppliers.

There is also a practical reason this approach keeps coming back. AI demand remains high, and compute shortages have not gone away. When capacity is tight, the best-performing chip is not always the one with the highest raw throughput. Often it is the one that can deliver the most useful work for the least power, with the fewest added infrastructure demands.

For now, the Google Frozen v2 AI chip remains a reported project rather than a confirmed product. But the outline of the strategy is clear enough. Google seems to be preparing for a future where Gemini inference, voice workloads, and power efficiency all matter at once, and where cutting NVIDIA reliance is part of a longer-term hardware plan.