Alphabet Develops Frozen v2 Chip to Cut Gemini Inference Costs

Key Takeaways

• Alphabet is developing Frozen v2, a custom inference chip that could enter service as early as 2028.
• The chip reportedly embeds elements of Gemini in silicon and could deliver 6 to 10 times the efficiency of Alphabet’s current AI chips.
• Lower inference costs could strengthen Google Cloud, protect Google search margins, and intensify competition with NVIDIA, AMD, and Intel.

Alphabet is developing a custom AI chip called Frozen v2, a project intended to reduce the computing and energy required to run Gemini models. The chip reportedly embeds architectural elements of Gemini directly into physical silicon, limiting how frequently data needs to move between processors and memory. Alphabet could begin using Frozen v2 as early as 2028. If the reported efficiency target holds, the chip could operate 6 to 10 times more efficiently than Alphabet’s current AI processors.

That matters because the economics of generative AI increasingly hinge on inference, the process of generating an answer after a model has been trained. Training is expensive but periodic. Inference happens every time Gemini summarizes a document, produces an image, answers a search query, or supports an enterprise application. At Alphabet’s scale, small savings on each request can add up quickly across billions of interactions.

Model quality is only one competitive variable. AI providers also need to manage electricity, memory, networking, cooling, and data-center capacity. Engineers commonly measure that burden through cost per token and inference efficiency. A token is the smallest unit of text processed by a model, typically around four characters. What happens when two models offer comparable performance, but one costs substantially less to operate? The lower-cost system gains room to reduce prices, serve more users, or improve margins.

The semiconductor market is already reflecting this shift. Gartner forecast worldwide AI semiconductor revenue of about $71 billion in 2024, up 33% from 2023, followed by nearly $92 billion in 2025. Data-center semiconductor spending also nearly doubled in 2024, rising from $64.8 billion to $112 billion. Meanwhile, IDC projected that AI infrastructure semiconductor revenue could increase from $42.1 billion in 2022 to $193.3 billion by 2027. CIO Dive has likewise documented how AI demand is reshaping chip investment and competition.

Frozen v2 gives Alphabet a potential advantage because it controls several layers of the AI stack. Alphabet develops Gemini, designs custom accelerators, operates global data centers, sells capacity through Google Cloud, and distributes AI through Google search and other consumer products. NVIDIA, AMD, and Intel compete in AI accelerators, but Alphabet can optimize hardware around its own models and workloads. That tighter connection can help reduce overhead, although embedding model-specific elements in silicon may create trade-offs when Gemini’s architecture changes.

The financial capacity behind the project is a crucial advantage. Alphabet expects capital expenditures and operating cash flow to remain at levels few AI developers can match, though specific budget targets for Frozen v2 are not disclosed. OpenAI, Anthropic, and xAI have collectively absorbed billions of dollars in infrastructure costs while pursuing increasingly capable models. Many startups therefore rent computing capacity rather than build data centers themselves.

Google Cloud is central to the strategy. The segment generates a critical portion of Alphabet’s sales and competes heavily with the cloud businesses of Amazon and Microsoft (source). Gemini now has 950 million monthly active users, while Alphabet’s models process 22 billion tokens per minute. Anthropic has also committed to substantial Google Cloud spending over five years.

Advertising remains the larger prize, accounting for the vast majority of Alphabet’s sales last year. AI Overviews and Gemini-powered search features can answer questions and compare products directly, but they also introduce additional inference costs into a business built around enormous query volume. Frozen v2 could help Alphabet absorb those costs while preserving advertising economics. The AI contest is becoming increasingly operational, and Alphabet is betting that cheaper intelligence will matter as much as smarter intelligence.