Google isn’t just building AI models anymore. It’s building the silicon to run them.
Alphabet is developing a custom server chip codenamed “Frozen v2” specifically optimized for its Gemini AI models, with a target launch date of 2028. The chip is designed to generate 6 to 10 times more tokens per unit of power than Google’s latest Tensor Processing Units.
Investors noticed. Alphabet shares climbed approximately 3% following the report, a clear signal that Wall Street sees custom silicon as more than a vanity project.
The economics of rolling your own chips
Frozen v2 is what the industry calls an ASIC, an application-specific integrated circuit. Think of it like this: Nvidia’s GPUs are Swiss Army knives, versatile and capable of handling lots of different tasks. An ASIC is a scalpel, designed to do one thing extraordinarily well. In Google’s case, that one thing is running Gemini inference workloads as cheaply and quickly as possible.
And Google isn’t exactly starting from scratch here. The company has been designing its own TPUs since 2015, making it one of the earliest major tech players to invest in custom AI silicon. Frozen v2 represents the next evolution of that strategy, moving from general-purpose AI acceleration to chips tailored for a specific model family.
A growing club of chip designers
Google isn’t alone in this pursuit. OpenAI and Anthropic are also developing bespoke AI chips, all driven by the same fundamental pressure: Nvidia’s GPUs are expensive, supply-constrained, and not always optimized for the specific workloads these companies care about most.
The custom AI chip market is projected to grow approximately 45% year-over-year in 2026, reflecting a broader industry shift away from reliance on a single hardware vendor.
What this means for investors
Alphabet’s $180 to $190 billion capital expenditure guidance is focused on AI infrastructure. A custom chip that dramatically cuts inference costs is exactly the kind of evidence that justifies massive capex.
The risk side of the equation is worth noting too. Custom chip development is notoriously expensive and slow. A 2028 target launch means Frozen v2 needs to outperform whatever Nvidia, AMD, and other chipmakers ship between now and then.
Still, Google’s track record with TPUs suggests this isn’t a speculative moonshot. The company has successfully deployed multiple generations of custom AI hardware at scale, including the seventh-generation TPU named “Ironwood” in late 2025.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
