Alphabet is reportedly working on a custom server chip, internally called “Frozen v2,” intended to make its in-house Gemini models run far more efficiently. The Information first reported details about the project, saying the new processor could deliver between six and ten times more tokens per unit of power than Google’s current AI chips, and that the company is aiming for a 2028 deployment.
What the reports say about Frozen v2
According to the reporting, Frozen v2 is being developed to optimize inference workloads for Google’s Gemini family of large language models. The claimed efficiency improvement — six to tenfold gains in tokens per watt — would represent a major step in reducing the energy and cost required to serve large-scale generative AI.
Google did not directly confirm the report when approached by media. In public comments, the company said its teams continuously explore hardware and software innovations to improve performance and efficiency, noting that not every experiment moves into production. That response underscores Google’s established approach of co-designing hardware and software to align systems with real-world workloads.
Why custom chips matter for AI providers
Building proprietary AI processors has become a strategic priority for prominent AI developers. Custom chips can be tailored to the specific computational patterns of a company’s models, improving energy efficiency, latency, and throughput. These gains matter because running large language models at scale is expensive and power-hungry, and efficiency improvements directly affect operating costs.
Beyond cost, vertical integration around hardware gives firms more control over supply and product roadmaps. Recent shortages and high demand for AI compute have prompted major cloud and AI companies to seek alternatives to off-the-shelf datacenter accelerators.
Industry context: peers and competition
Google is not alone in pursuing in-house silicon. OpenAI disclosed its first custom inference processor, named Jalapeño, earlier this year. Anthropic has reportedly been in talks with Samsung on chipmaking collaborations. These moves reflect a broader effort by AI companies to reduce dependence on established chip vendors, most notably Nvidia, which has dominated the AI accelerator market.
Nvidia’s GPUs have powered much of the recent surge in model training and inference, but their dominance has motivated large AI companies to explore alternatives that can be better tuned to their software stacks and business needs. Custom chips also create an opportunity to differentiate on cost and performance for cloud customers and end-users.
Financial and strategic implications for Alphabet
News of the Frozen v2 reports arrived against the backdrop of substantial capital commitments from Alphabet to expand its AI capabilities. Earlier this year, Google disclosed plans to spend between $180 billion and $190 billion, investments that include datacenter expansions, networking, and AI infrastructure. Those outlays prompted scrutiny from investors keen to see returns on such scale.
Market reaction to the Frozen v2 report appeared positive: after the story was published, Alphabet’s stock rose roughly 3% on the following Monday morning. The move suggests investors responded to the possibility that more-efficient hardware could improve margins and accelerate the company’s ability to deploy demanding AI services at scale.
Practical hurdles and timeline
Developing a competitive AI datacenter chip is technically demanding and costly. The reported 2028 timeframe for Frozen v2 indicates a multi-year effort — consistent with the timelines typically required for architecture design, fabrication partnership negotiations, and integration into server platforms. Even when prototypes exist, companies often iterate through multiple silicon generations before reaching wide production.
Google’s previous hardware initiatives, such as its Tensor Processing Units (TPUs), demonstrate both the potential benefits and the long lead times of custom silicon programs. The company’s emphasis on co-designing hardware and software helps ensure that any new chip would be optimized for Google’s datacenter stack and model workloads.
What to watch next
Key developments to follow include any formal confirmation from Alphabet, technical disclosures about Frozen v2’s architecture or manufacturing partners, and signs of integration with Google Cloud or internal Gemini production environments. Observers will also look for regulatory filings or capital commitments tied to chip development and fabs, which could reveal more concrete timelines and costs.
For the broader market, further moves by major AI companies to internalize chip design could reshape supplier dynamics and competitive positioning. Efficiency gains at the hardware level would also feed into broader discussions about AI operating costs, sustainability, and the economics of large-scale generative services.
For now, Frozen v2 remains an unconfirmed yet notable element of Alphabet’s AI strategy, emblematic of an industry-wide push toward specialized silicon designed to match the demands of modern AI models.
Source: TechCrunch AI