Alphabet builds dedicated Gemini chip to cut compute costs
Alphabet is developing a specialized chip for its Gemini model that could drastically reduce operating costs and alleviate the server capacity constraints currently forcing the company to turn away customers.
Alphabet is developing a specialized server chip designed exclusively to run its Gemini AI model, targeting deployment by 2028. Codenamed Frozen v2, the silicon departs from Google’s traditional Tensor Processing Units by hardwiring Gemini’s architectural blueprint directly into the hardware. Engineers project this approach will yield six to ten times the efficiency in tokens generated per watt of electricity compared to current TPUs.
The initiative addresses an acute capacity bottleneck. Despite committing up to $190 billion to AI infrastructure this year, Google recently informed Meta it could not fulfill the compute volume requested, forcing the social media giant to ration employee AI usage. Purpose-built hardware offers a potential path to serve surging demand without proportionally exploding power consumption and capital expenditures.
Cost efficiency is rapidly becoming a defining factor in the commercial AI market. Rivals including OpenAI, Anthropic, and Chinese labs currently capture up to 45% of AI token usage by U.S. companies. They achieve this largely by operating at 60% to 90% lower costs than U.S. hyperscalers. A cheaper-to-run Gemini directly protects Alphabet’s margins against this price pressure. Because Frozen v2 is hardwired for one model, it will not be offered to external cloud customers.
Investors initially cheered the strategic shift, pushing Alphabet shares up roughly 3% to touch $356 intraday on Monday. Those gains faded by Tuesday as the market awaited the company’s second-quarter 2026 earnings report scheduled for Wednesday, July 22. The interim reality of Google's hardware constraints was underscored by reports that the company is paying SpaceX $920 million a month to rent 110,000 Nvidia GPUs from xAI data centers.
The project highlights a sustained, industry-wide effort to loosen Nvidia’s grip on the AI supply chain. Nvidia controls roughly 85% of the AI GPU market using hardware originally engineered for video games, which carries inherent overhead when processing language models. Alongside Amazon, Microsoft, and Meta, Google is betting that application-specific silicon will eliminate this inefficiency and translate into billions in long-term savings. However, Frozen v2 remains exploratory, with key design decisions still unresolved and no official confirmation from the company.
