Google developing a Gemini-specific chip, "Frozen v2," claimed 6-10x more efficient than current TPUs

Google is working on a new chip called Frozen v2, designed specifically to run Gemini models more efficiently. The Information reports the chip could be 6 to 10 times more efficient than the TPUs Google uses today. The target date is 2028, and it’s meant to complement TPUs, not replace them.

The efficiency claim matters because running large AI models is expensive. If you’re using Gemini through an API or building it into your products, lower costs per query could mean cheaper bills or faster responses at the same price. Google would also gain an edge over competitors who rely on Nvidia chips or their own custom silicon. Alphabet’s stock moved on the news, which tells you investors think this could shift economics in a meaningful way.

But 2028 is a long time in AI. A lot can change between now and then, including what models look like, what tasks matter most, and whether the efficiency gains hold up in practice. The report doesn’t say how much Frozen v2 will cost to manufacture or deploy, which is half the equation. Still, the direction is clear. Google is betting that purpose-built chips for specific models will outperform general-purpose accelerators, at least for their own stack.

For business owners, this is a reminder that the cost structure of AI is still moving. If you’re planning multi-year contracts or building infrastructure around current pricing, keep an eye on these hardware shifts. They won’t hit overnight, but they could change what’s affordable or practical by the time your next renewal comes up. This is the kind of trend we track in the Omni Command Centre, where you can see how infrastructure changes might affect your AI roadmap before they show up in your invoices.