Google reportedly building a chip with Gemini's architecture etched into silicon

Google is developing a server chip, internally called "Frozen v2," that embeds the architecture of its Gemini AI models directly into silicon rather than relying on general-purpose tensor processing, according to a report from The Information citing two people with direct knowledge of the project. The design reportedly traces back to an idea from Jeff Dean, Google DeepMind's chief scientist, and could deliver 6 to 10 times more AI tokens per watt than Google's current-generation TPUs.

Google has not publicly confirmed Frozen v2's existence, though the company has acknowledged ongoing hardware research broadly. The report is significant regardless of official confirmation because it reflects a structural shift in how the largest AI labs are approaching inference costs: rather than building increasingly general hardware and hoping software catches up, Google is reportedly willing to hardwire a specific model's architecture into a chip that can only run that architecture's future versions.

What "architecture etched into silicon" actually means

A standard TPU or GPU is a general-purpose accelerator: it can run any model architecture a developer feeds it, at the cost of spending computation on steps that a fixed pipeline could handle more efficiently. Frozen v2 reportedly takes a middle path between that flexibility and a fully fixed-function chip. Rather than freezing the model's trained weights — an earlier concept that Google apparently abandoned because weights go stale too quickly to justify custom silicon — Frozen v2 embeds Gemini's architectural blueprint: the sequence of operations, routing decisions, and structural choices that define how the model processes a query, independent of the specific parameter values.

New weights can still be loaded onto the chip as Gemini is retrained and updated. What's fixed in the transistors is the shape of the computation itself, which reduces the number of steps the chip has to take and the amount of data it has to move per query — the two factors that dominate inference energy cost at scale.

The tradeoff: efficiency for flexibility

The cost of that efficiency gain is architectural lock-in. Because Gemini's structure is physically built into the chip, Frozen v2 can only support future Gemini versions as long as Google preserves the model's foundational architecture. If Google's research teams make a fundamental change to how Gemini is structured — the kind of shift that happens periodically across model generations — the chip stops being useful for anything beyond its original purpose.

That's almost certainly why reports describe Frozen v2 as an internal-use chip rather than a future commercial product in the mold of Google's TPU line, which Google rents out via Cloud and licenses more broadly. A chip locked to one company's proprietary model architecture has no market outside that company.

Timeline and strategic context

Google is reportedly targeting deployment starting in 2028, with production volumes smaller than its mainline TPU program — consistent with Frozen v2 functioning as a proof of concept for specialized, architecture-specific silicon rather than a broad infrastructure bet. The project fits a wider 2026 pattern: AI labs and cloud providers are increasingly treating inference cost, not training cost, as the binding constraint on how widely they can deploy their models, and custom silicon is one of the few remaining levers once model efficiency improvements from software alone start to plateau.

If the 6-to-10x efficiency figures hold in production, Frozen v2 would meaningfully reduce Google's per-query compute cost for Gemini specifically — a direct answer to the compute capacity constraints that have shaped Google's AI roadmap and pricing throughout 2026. As reported by The Information and corroborated by multiple technology outlets, the project underscores that architectural specialization, not just raw compute scale, is becoming a competitive axis among frontier AI labs.