Alphabet is developing a new AI server chip, internally known as Frozen v2, that would embed elements of its Gemini model directly into the hardware itself. The project, reported by The Information and confirmed in substance by multiple outlets, represents a meaningfully deeper commitment to custom silicon than Google’s existing Tensor Processing Units already reflect.
Where TPUs remain general-purpose accelerators capable of running a range of machine learning workloads, Frozen v2 is reportedly optimized specifically for Gemini’s architecture. Portions of the model would be permanently hardwired into the chip rather than left flexible in software, cutting down on the calculations and data movement required every time the system processes a query. Google engineers cited in the report project the chip could deliver six to ten times more AI tokens per unit of power than the company’s newest TPUs.
The efficiency framing matters, but the more consequential detail is the capacity problem sitting underneath it. Frozen v2 is reportedly aimed at easing an internal AI compute shortage that has grown severe enough for Google Cloud to decline external customer deals. A hyperscaler turning away paying cloud business because it needs the capacity for its own models is a stronger signal of constraint than any efficiency multiple.
Alphabet shares responded accordingly, closing roughly 1.5% higher on the report, with intraday gains as high as approximately 3.7% as investors read the news as a credible new lever against rising AI infrastructure costs heading into earnings.
The trade-off is architectural lock-in. Because parts of Gemini are baked directly into the silicon, Frozen v2 would only remain useful for future Gemini generations if Google preserves the same underlying architecture. That is a materially larger bet than a routine TPU refresh, and it signals Google now views Gemini’s architecture as stable enough to commit to hardware permanence, sacrificing the flexibility that general-purpose chips retain.
Scope should not be overstated. Google reportedly treats Frozen v2 as a new, complementary chip family rather than a TPU replacement, with a considerably smaller initial production run that functions partly as a proving ground for increasingly specialized AI hardware as foundation models mature. Deployment is targeted for around 2028, which places this firmly outside any near-term supply catalyst.
The strategic context is the broader push by hyperscalers to reduce dependence on Nvidia’s merchant GPUs through custom silicon, a trend already visible in Broadcom’s TPU manufacturing partnership and Marvell’s custom ASIC design business. Frozen v2 extends that trajectory a step further by moving from general-purpose custom silicon toward model-specific silicon. It also reinforces the case for advanced packaging capacity, since increasingly specialized AI chips depend on tighter integration between compute and memory. That keeps the hybrid bonding and HBM packaging theme relevant even where the headline is about a single hyperscaler’s internal chip roadmap rather than the merchant memory or foundry supply chain directly.
Google has not confirmed Frozen v2 by name. In response to inquiries, the company said its teams continuously research new hardware and software approaches to improve performance and efficiency, without denying the specific project.
