Google Building Gemini-Native Server Chip Called Frozen
The Neuron ● Covered by 19 sources
Google's reportedly building a new server chip called Frozen, made to run its Gemini AI models natively. It's another sign that the big AI labs want their own custom silicon instead of Nvidia's.
Google has apparently gone deeper into custom chip territory with a project called Frozen, a server-grade chip built specifically to run Gemini models. The name sounds like something out of a spy thriller, but the intent is straightforward: Google wants hardware that speaks Gemini's language natively, rather than hardware that has to be coaxed into it through layers of software translation.
This isn't Google's first rodeo with custom silicon. The company has been building TPUs, its Tensor Processing Units, for years, and those chips already power a huge chunk of Google's AI training and inference workloads. Frozen appears to be a further step in that direction, tailored tightly to Gemini rather than being a general-purpose accelerator that happens to run Gemini well.
The timing lines up with a broader shift happening across the industry. Nvidia's GPUs have been the default choice for AI compute for years, and that dominance has made Nvidia one of the most valuable companies on the planet. But the costs and supply constraints of relying on someone else's chips have pushed Amazon, Microsoft, and Meta to all pursue their own custom silicon too, each betting that owning the hardware stack gives them more control over cost, speed, and how tightly software and chip can be fused together.
For Google specifically, a Gemini-native chip could mean serving its models faster and cheaper at scale, which matters enormously as it pushes Gemini into search, Workspace, Android, and pretty much every product surface it has. Less reliance on Nvidia also means less exposure to that company's pricing power and chip shortages, something every hyperscaler has felt sting at some point in the last two years.
Details on Frozen's specs, timeline, or manufacturing partner haven't surfaced yet, and Google hasn't confirmed the project publicly. But the direction is clear enough: the biggest AI companies increasingly want chips built around their models, not the other way around.
My take
I'll say what everyone in the open-model camp is thinking: this is exactly the kind of vertical lock-in that should worry people who care about competition in AI. When a company controls the model, the chip, the cloud, and the distribution, you don't get an ecosystem, you get a walled garden with really good marketing. I'd rather see more investment in open silicon standards than another Big Tech chip built to make switching providers harder.
Read more about this at: The Neuron
