Everyone thinks winning the AI race is all about building the smartest computer program. Google seems to be betting on something different, the hardware humming quietly underneath it all.
For years, Google has quietly spent huge amounts of money building its own computer chips. Nobody writes big news stories about this kind of work. It’s slow. It’s not flashy. Most people never even hear about it.
But it ends up deciding how fast AI feels to us, and how cheaply and efficiently it can run at scale. Now a new report from The Information is giving us a peek behind the curtain.
According to a report from The Information, Google is developing another custom chip, internally nicknamed Frozen v2. It’s made specifically to help Gemini run faster while using a lot less power.
Google hasn’t confirmed the details or a launch date. But the company has said it always builds hardware and software together to get the best performance. If this report turns out to be true, this chip could become a big deal for Google’s future plans.
What Is Google’s New AI Chip
According to recent reports, the project is internally known as Frozen v2 and is being developed for one specific task: running Gemini inference more efficiently. Inference is the stage where an AI model processes a prompt and generates a response.
It might feel instant to you and me, but serving billions of these requests every single day is one of the most compute intensive and expensive parts of running modern AI services.
How Frozen v2 Is Reportedly Designed
Most chips can run any program. Frozen v2 is different. That’s because reports say Gemini’s design is built right into the chip.
So why does that matter? Building Gemini’s design into the chip cuts out wasted work. Here’s the key part though. The weights, or settings, can still change. But the structure stays locked in, or “frozen.” And since the chip is made for just one job, it ends up using a lot less power.
So how big is the gain? Insiders say it could be six to ten times better than Google’s chips today. That number is about power use, though, not speed. Google hasn’t confirmed it yet.
As for timing, Google wants 2028. But that could still change, since the design isn’t done yet.
Why Google Is Treating This as a Trial Run
Here’s an interesting twist. Reports say Google isn’t fully committing to Frozen v2 yet, it’s more of a test run.
Google reportedly won’t build as many of these chips as it builds TPUs, its own AI chips already running most of its models. Frozen v2 doesn’t replace them; it’s meant to work alongside TPUs, handling one job: running Gemini.
The catch is flexibility. Since parts of Gemini are built right into the chip, it only stays useful if Gemini doesn’t change too much. If Google reworks Gemini later, parts of this chip could stop working well.
Why Efficiency Suddenly Matters So Much
Building a smart AI model is only half the battle. The real cost shows up when you run it for millions of people every day.
Here’s what’s really going on. Gemini uses power, memory, and electricity every time it answers a question. Multiply that by billions of daily requests, and costs pile up quickly.
Reports suggest Google has hit some real limits on computing power, bad enough that Google Cloud has reportedly had to turn away some outside work.
So what would Frozen v2 actually fix? If those efficiency gains hold up, Google could handle more requests while burning less power, which means faster responses, lower costs, and getting more out of what it already owns.
Potential benefits of Frozen v2 include:
- Faster Gemini responses
- Lower electricity use across Google’s data centers
- More computing room for enterprise customers
- Better use of the AI infrastructure Google already has
- Lower operating costs over time
Why This Reaches Beyond Google
Google isn’t the only company investing in specialized AI hardware. Across the industry, major AI developers are looking beyond bigger models and turning to custom silicon instead, chasing lower infrastructure costs, better efficiency, and more control over how their AI systems run at scale.
More companies are building this same kind of chip now. It’s a bigger trend across AI. Instead of buying regular chips, companies are building chips made just for their own AI models.
Why does this matter? It just works better. It costs less too. And it lets companies control how their hardware and software fit together.
The Bigger Picture Right Now
None of this is happening by accident. Competition in AI keeps heating up. Because of that, it’s not just about who builds the biggest model anymore.
So what are companies doing instead? They’re spending big money on special hardware too. They want better results while spending less. That’s where Frozen v2 comes in.
If the reports are true, this chip is Google’s way of making Gemini and Google Cloud stronger and more reliable. And that’s a big deal, since good hardware is starting to matter just as much as the AI models themselves.
What This Means for Gemini Users
If Frozen v2 eventually makes it to production, most of what changes will happen quietly behind the scenes. Businesses running things through Gemini on Google Cloud could see better scalability as demand keeps growing.
Everyday users might notice faster responses, lower lag, and steadier performance during busy stretches. Most users will never know the chip exists. They’ll simply notice that Gemini feels faster, more responsive, and more reliable over time.
Wrap Up
Nobody knows yet if Frozen v2 will actually turn out exactly as reported. But either way, this story shows something important happening across AI right now. Building a smart AI model just isn’t enough anymore.
Companies also need to run those models well, keep them reliable, and make them work at a huge scale. If Google gets this right, Frozen v2 could end up being one more important piece behind the next generation of Gemini.
