Alphabet, Google's parent company, is designing a new server chip internally dubbed 'Frozen v2' to help its in-house Gemini AI models operate more efficiently. The chip is slated for release sometime in 2028, according to a report citing anonymous sources familiar with the project. The report claims that Frozen v2 could be between six and ten times more efficient than Google's existing AI chips, when measured by the number of tokens generated per unit of power.
Google's Custom Silicon Journey
Google has a long history of developing custom hardware for its AI workloads. The company introduced its first Tensor Processing Unit (TPU) in 2016, a custom application-specific integrated circuit (ASIC) specifically designed to accelerate machine learning tasks. Since then, Google has released multiple generations of TPUs, with the latest being the TPU v5p and the more specialized TPU v5e for inference. These chips power Google's own services like Search, Translate, and Gmail, as well as its cloud customers through Google Cloud Platform. However, as AI models grow larger and more complex, the demand for more efficient hardware has never been higher. The Frozen v2 project appears to be a departure from the TPU lineage, focusing specifically on the needs of Gemini, Google's most advanced multimodal AI model family.
Why Efficiency Matters
The efficiency of an AI chip is measured in tokens per watt – the number of output tokens a model can generate for each unit of electrical energy consumed. A chip that is six to ten times more efficient could dramatically reduce the operational costs of running Gemini at scale. This is particularly important as companies like Google spend tens of billions of dollars annually on AI infrastructure. According to Alphabet's recent quarterly filings, capital expenditures for 2025 were projected between $180 billion and $190 billion, with a significant portion allocated to AI compute resources. Improving chip efficiency directly impacts the bottom line, making AI services more cost-effective and potentially more accessible to smaller businesses and developers.
Efficiency also addresses growing environmental concerns. Data centers that run AI workloads consume massive amounts of electricity – some estimates suggest that AI could account for 4% of global electricity consumption by 2030. By making chips that do more per watt, Google can reduce its carbon footprint while maintaining or increasing its AI capabilities. The company has set ambitious sustainability goals, including operating on 24/7 carbon-free energy by 2030, and more efficient silicon is a key part of that strategy.
The Race for Custom AI Silicon
Google is not alone in pursuing custom chips for AI. Over the past year, several major AI companies have announced their own hardware initiatives. In June, OpenAI unveiled its first custom inference processor, named 'Jalapeño,' designed to optimize the performance of its GPT models. Earlier this month, reports emerged that Anthropic, another leading AI research firm, was in discussions with Samsung for a new chipmaking partnership. These moves reflect a broader industry trend: AI model providers are increasingly seeking to reduce their dependence on Nvidia, which has dominated the AI chip market for years with its GPU-based accelerators.
Nvidia's GPUs, particularly the H100 and B200 Blackwell, are the gold standard for training and running large-scale AI models. However, the company's market dominance has led to high prices, long lead times, and supply constraints. By developing their own chips, companies like Google, OpenAI, and Anthropic can optimize hardware specifically for their models, potentially achieving better performance per dollar and per watt. They also gain more control over their supply chains and can avoid the vendor lock-in that has worried many in the industry.
Competitive Landscape and Investor Sentiment
The news of Frozen v2 appears to have boosted investor confidence in Alphabet. Following the chip-related report, the company's stock rose approximately 3% on a Monday morning, ahead of its quarterly earnings report later that week. This uptick suggests that investors are optimistic about Google's ability to control costs and innovate in hardware amid a backdrop of intense capital spending. The broader AI market has experienced a recalibration over the past year, with investors shifting from euphoric growth expectations to a more sober assessment of profitability and efficiency. Companies that can demonstrate a path to lower operational costs are likely to be rewarded by the market.
Alphabet's massive planned expenditures – between $180 billion and $190 billion – have previously raised concerns among shareholders. The company has been investing heavily in data center construction, networking infrastructure, and chip development. If Frozen v2 delivers on its promised efficiency gains, it could help justify those outlays by reducing the total cost of ownership for Google's AI infrastructure. Additionally, the chip could become a product offered to Google Cloud customers, creating a new revenue stream outside of the company's core advertising business.
What Frozen v2 Means for Gemini
Gemini is Google's flagship AI model family, designed to compete with OpenAI's GPT-4 and Anthropic's Claude. It powers numerous Google products, including the Gemini chatbot, Google Workspace features, and various cloud APIs. The model is multimodal, meaning it can understand and generate text, images, audio, video, and code. Running such a complex model at scale requires immense computational resources. A chip that is significantly more efficient could enable Google to deploy Gemini with lower latency, higher throughput, and reduced energy consumption. This might allow the company to offer more competitive pricing for its AI services and expand Gemini's capabilities without proportionally increasing its carbon footprint.
The timeline of 2028 suggests that Frozen v2 is a long-term project. In the interim, Google will continue to rely on its TPU lineup and potentially on chips from Nvidia and other suppliers. However, the development of a custom chip specifically for Gemini indicates that Google sees value in vertical integration – designing hardware and software together for optimal performance. This full-stack approach is something Google has long emphasized, most notably in its Pixel phones with the Tensor chip, and now it is extending the same philosophy to its data center AI workloads.
Challenges and Technical Hurdles
Developing a custom AI chip is not without risks. The semiconductor industry is notoriously capital-intensive and complex, with long design cycles, manufacturing constraints, and fierce competition. Google's previous TPU generations have been successful, but the company has also experimented with other custom silicon projects – like the video codec chips and Titan security chips – that have not always been widely adopted. Moreover, the AI landscape is evolving rapidly; by the time Frozen v2 arrives in 2028, the requirements for AI models may have shifted. Models could require even more memory bandwidth, different numerical precision formats, or entirely new architectures like sparse attention mechanisms that are not well-served by traditional ASIC designs.
Another challenge is manufacturing. Google does not own its own fabrication plants (fabs); it relies on contract chipmakers like TSMC or Samsung to produce its designs. The foundry industry is subject to geopolitical risks, supply chain disruptions, and technology node limitations. As chips become smaller and more complex, manufacturing costs rise and yields can be unpredictable. Google must also secure enough production capacity to meet its internal demand, especially if Frozen v2 is also offered to external cloud customers.
Broader Industry Implications
The push for custom AI chips by major tech companies could reshape the semiconductor supply chain. If Google, OpenAI, Anthropic, and others succeed in producing efficient alternatives to Nvidia's GPUs, it could erode Nvidia's dominant market share over time. However, Nvidia is itself innovating rapidly, with its next-generation Rubin architecture expected in 2026, so the competitive dynamics are far from settled. Additionally, startups like Cerebras, Graphcore, and Groq are developing novel AI chip architectures that challenge both Nvidia and the hyperscaler custom chips.
For the average user of Google services, the benefits of Frozen v2 may not be immediately apparent. But over time, more efficient AI hardware could lead to faster responses, better-quality outputs, and lower costs for Google's free and paid AI offerings. It could also enable new AI applications that were previously too expensive or power-hungry to run at scale. The ripple effects will likely be felt across the entire tech ecosystem, from cloud computing to consumer devices.
The Frozen v2 project also highlights the growing importance of hardware-software co-design. Google has heavily invested in its own TensorFlow and JAX frameworks, which are used to train and deploy models on its TPUs. By designing a chip specifically for Gemini, Google can ensure that the hardware and software are tightly integrated, avoiding the performance inefficiencies that often arise when using generic accelerators. This co-design philosophy is something that Apple has long practiced with its A-series and M-series chips, and it is increasingly being adopted in the AI space.
As the AI industry matures, the winners will not just be those with the most advanced models, but those who can deploy them most efficiently. Google's Frozen v2 represents a bet that custom silicon will be a key differentiator in the years ahead. Whether the chip meets its ambitious efficiency targets and comes to market on schedule will have significant implications for Alphabet's AI ambitions and its financial performance. For now, the company's stock response suggests that investors are willing to give Google the benefit of the doubt – but the real test will come when Frozen v2 moves from the drawing board to the data center floor.
Source: TechCrunch News