Google's Frozen v2 Chip Hardwires Gemini for 10x Efficiency

Google is developing a specialized server chip, internally codenamed Frozen v2, to embed the architecture of its Gemini models directly into silicon. The ongoing hardware project marks a pivot from flexible accelerators toward hardwiring specific model components for high-volume inference. If you rely on the Gemini API at scale, this physical integration aims to drastically reduce the computational overhead required for every query.

Hardware-Level Model Integration

General-purpose accelerators and Google’s own Tensor Processing Units (TPUs) are designed to support a wide variety of model architectures. Frozen v2 abandons this flexibility to incorporate specific elements of the Gemini architecture directly into the chip’s circuitry. This minimizes data movement and heavily optimizes the exact mathematical operations Gemini requires.

Google engineers are currently finalizing the design for a targeted deployment in 2028. The goal is to achieve a 6x to 10x improvement in power efficiency, measured by the number of AI tokens served per unit of power consumed.

Hardware	Primary Function	Architecture Flexibility	Target Workload
TPU v7 (Ironwood)	Training and Inference	High	Broad general AI workloads
TPU v8 (Sunfish/Zebrafish)	Training and Inference	High	Broad general AI workloads
Frozen v2	Inference Only	Low (Gemini-specific)	High-volume Gemini queries

Frozen v2 will operate alongside the existing TPU lineup rather than replacing it. Flexible chips like the recently released TPU v7, which provides 4.6 PetaFLOPS of FP8 performance, will continue handling broader AI inference tasks and model training.

Infrastructure Constraints

The development of Frozen v2 stems from a severe internal shortage of AI computing power. This capacity crunch has reportedly forced Google Cloud to turn away certain enterprise customers and prompted the company to secure temporary compute capacity from SpaceX.

Alphabet recently guided $175 billion to $185 billion in capital spending for 2026, heavily concentrated on scaling infrastructure. The Frozen v2 project aligns with the full-stack strategy highlighted by CEO Sundar Pichai at Google I/O 2026 in May. By controlling the entire pipeline from custom silicon to the application layer, Google can optimize performance and cost structures tightly.

The broader market is moving in a similar direction. OpenAI recently announced its own inference chip, Jalapeño, developed with Broadcom in June 2026. Anthropic is also reportedly negotiating with Samsung for custom AI silicon to run its models.

As models scale, raw compute limits dictate API availability and latency. Dedicated ASIC designs like Frozen v2 trade future architectural flexibility for immediate, massive efficiency gains on current flagship models. Developers building on the Gemini platform should monitor these hardware shifts, as they directly influence future pricing tiers, speed limits, and context window constraints.