Google is reportedly building a Gemini-only chip
Frozen v2 etches Gemini's architecture into silicon for a claimed 6-10x efficiency gain over Google's current TPUs, with rollout targeted for 2028.
Google is reportedly developing a server chip called Frozen v2 that hardwires Gemini’s model architecture directly into the silicon, according to sources cited by The Information and corroborated by multiple outlets. The pitch: by fixing Gemini’s structural decisions in the transistors rather than running them through general-purpose hardware, the chip could deliver 6 to 10 times better performance per watt than Google’s current TPUs, through fewer computation steps and less data movement between on-chip and off-chip memory.
Frozen v2 is a scaled-back version of an earlier “Frozen” concept from Google DeepMind chief scientist Jeff Dean, which would have baked Gemini’s actual weights into the chip. Google reportedly shelved that approach because hardware tied to one specific model version would go stale too fast, a risk that looks bigger given how fast Gemini itself has been moving, including the ground-up Gemini 3.5 Pro rebuild that shipped this month. Frozen v2’s compromise freezes the architecture while keeping weights updatable, and Google plans to deploy it starting in 2028, running alongside rather than replacing its TPU line and Google’s existing optical-switch cluster infrastructure. It won’t be sold externally.
None of this is a Google on-the-record announcement, so the efficiency numbers and 2028 timeline are projections from people familiar with an internal project, not measured results. The context explains the motivation regardless: Google has reportedly been throttling some external cloud workloads to keep up with Gemini demand, and inference efficiency increasingly sets the ceiling on how aggressively a lab can price its API against competitors. If Frozen v2 ships anywhere near its claimed numbers, it’s a lever for Google to undercut OpenAI and Anthropic on Gemini API pricing later this decade, the same cost pressure already pushing Microsoft to shift Copilot traffic onto its own models, worth tracking even at a two-year-plus horizon, but not yet a reason to change any near-term infrastructure decision.