Google's 'Frozen v2' AI Chip for Gemini

Frozen v2's core bet is baking Gemini's neural architecture directly into silicon, cutting the distance data travels and the calculations needed to answer a query. Google's own engineers project 6 to 10 times more tokens per unit of power than its latest TPUs ^[1]- a leap that would meaningfully lower the cost of serving Gemini at scale. But the approach uses a 'flexible hardwiring' method: it locks in the model's architecture, not its weights, so Gemini's parameters can still update after the chip ships ^[2]. That distinction matters because the same specialization that produces the efficiency gain is also the chip's biggest liability - if Google materially changes Gemini's underlying architecture, a Frozen chip could stop being useful, which is likely why the project is reportedly viewed partly as a trial run rather than a TPU-scale production line ^[1].

Despite the market excitement, this is explicitly not framed internally as a wholesale hardware shift. The 'Frozen' project is designed to create a new proprietary chip line that sits alongside Google's existing TPUs rather than supplanting them ^[3], and Google does not plan to produce Frozen v2 at the same scale as its TPU fleet ^[1]. Asked about the report, a Google Cloud spokesperson would not confirm specifics, instead pointing to the company's general practice of co-designing hardware and software from the ground up ^[1]. That carefully hedged response, paired with a 2028 target that leaves plenty of room for the design to change, underscores how much of this remains aspirational engineering rather than a settled product roadmap.

The timing points to real near-term pain. An AI computing capacity crunch has reportedly fueled internal tension at Google and pushed Google Cloud to decline deals with some outside customers ^[1]- a direct commercial cost of not having enough efficient inference capacity. That shortage sits alongside separate reporting that Google delayed its latest Gemini model launch after it missed internal goals, particularly on coding capability ^[1], adding urgency to an infrastructure bet that will not materialize for years. Because Frozen v2 is not slated to arrive until as soon as 2028, the capacity constraints motivating it will persist regardless of whether the chip ultimately succeeds ^[4].

Frozen v2 is Google's entry in a wider industry pattern of hyperscalers building proprietary AI accelerators to reduce dependence on Nvidia's costly, supply-constrained GPUs, alongside Microsoft's Maia chip and Amazon's Trainium processors ^[2]. Anthropic has also reportedly been exploring its own proprietary AI hardware ^[5]. Investors read the report as a sign Google is serious about controlling its AI infrastructure destiny: Alphabet shares rose roughly 3 to 3.5 percent the day the story broke ^[1], a move confirmed across multiple market reports ^[3][6], arriving just two days ahead of the company's Q2 2026 earnings.