Google’s Reported Frozen v2 Chip Could Make Gemini More Efficient

Google reportedly plans a Gemini-specific chip

Google is developing a new server chip that would bake parts of its Gemini AI model directly into silicon, according to a 20 July report by The Information. The project is reportedly called Frozen, with a second-generation design known as Frozen v2, and could be deployed as early as 2028.

The key idea is specialisation. Rather than relying solely on a broadly programmable AI accelerator to run Gemini’s calculations, Frozen v2 would hardwire elements of the model’s underlying architecture into the chip itself. That should reduce both the number of calculations required to answer a query and the amount of data that has to move around the system.

Engineers working on the project have projected that Frozen v2 could serve between six and ten times more tokens per unit of power than Google’s newest Tensor Processing Units, or TPUs, according to the report. Google has not publicly confirmed Frozen v2’s existence, its performance projections, or the eventual scope of its deployment.

Alphabet shares closed 1.51% higher on Monday following the report, which pointed to investor interest in hardware that could make Google’s fast-growing AI operation less expensive and less constrained by available computing capacity.

Frozen v2 would preserve flexibility where it counts

The name refers to the machine-learning practice of “freezing” parameters: locking values so they no longer change. But Frozen v2 reportedly takes a more measured approach than Google’s original proposal. It would freeze the architecture, rather than permanently embed Gemini’s model weights, leaving those weights updatable as Gemini evolves.

That distinction matters. A chip containing the weights of one specific Gemini release could become obsolete rapidly once Google updated its model. The earlier Frozen design, reportedly spearheaded by Google DeepMind chief scientist Jeff Dean, would have gone further by baking the weights into silicon. The Information reported that Google set that version aside because its useful lifespan would be too short.

Frozen v2 is intended to retain more room for software updates, but it would still be tied to the same fundamental Gemini architecture. If Google substantially changes that architecture in a future model generation, the specialised hardware may no longer fit the job. The report said the amount of Gemini to lock into the chip remains undecided.

That trade-off explains why Frozen v2 is reportedly being treated as an exploratory effort rather than a replacement for Google’s TPU programme. Production volumes are expected to be well below TPU levels.

A response to a shortage of AI compute

The reported project arrives while Google is trying to find enough computing power for its own products and Cloud customers. On Alphabet’s first-quarter 2026 earnings call, chief executive Sundar Pichai said the company was “compute-constrained in the near term”, adding that Cloud revenue would have been higher if Google had sufficient infrastructure to meet customer demand.

Google Cloud’s revenue rose 63% year-on-year to more than $20 billion in that quarter, while its backlog nearly doubled to more than $460 billion. The pressure is not limited to customer workloads: Google’s internal requirement for engineers to use AI for code generation also creates competition for the same GPU and accelerator resources.

The scale of the bottleneck is reflected in Google’s reported arrangements for external capacity. In June, Google agreed to pay SpaceX $920 million a month for access to roughly 110,000 Nvidia GPUs housed in xAI data centres, describing that capacity as a bridge for Gemini Enterprise demand it could not fully serve.

Google has also told Meta it could not provide the full amount of Gemini capacity Meta wanted to buy, according to the Financial Times. A highly efficient in-house inference chip would not solve every supply problem, but it could make Google’s own Gemini requests cheaper to run.

Frozen is separate from Google’s latest TPUs

Google only unveiled its eighth-generation TPUs at Cloud Next ’26 in April. The TPU 8t is designed for training, while the TPU 8i is intended for inference, reasoning and serving workloads. The latter has 384MB of on-chip SRAM, three times the amount in the previous Ironwood generation, to keep a larger KV Cache on the chip.

Frozen v2 is a parallel route: a purpose-built inference processor for one model family rather than a general platform for many AI workloads. That narrow purpose is also why Google Cloud customers are unlikely to be able to use it directly. A chip hardwired around Gemini would not be able to run customers’ own models.

What UK users should expect

There is no consumer product here, no UK price, and no confirmed launch timetable beyond the reported target of 2028 at the earliest. Frozen v2 would live in Google’s data centres, not in phones, laptops or home gadgets.

The eventual benefit for UK users of Gemini-powered Search, Workspace and Google One AI services could be faster responses or a wider rollout of AI features. But that is a possible downstream effect, not a promised outcome. For now, the significant development is Google’s willingness to explore turning a model architecture into hardware as inference becomes the dominant cost of operating AI at scale.