Alphabet has begun designing a custom server chip that could make its Gemini AI models dramatically more efficient, potentially easing a severe computing-capacity crunch inside Google Cloud. The project, known internally as “Frozen v2,” aims to process six to ten times more tokens per unit of power than Google’s latest tensor processing units (TPUs), according to a July 20 report by The Information. But the chip isn’t expected before 2028, and the company cautions that not every experimental design reaches production.
What’s actually being built
The Information’s report—confirmed by Alphabet in a statement to CNBC—details an accelerator that would permanently embed parts of the Gemini model architecture directly into silicon. Unlike general-purpose AI chips that must handle many model families, Frozen v2 is being tuned specifically for Google’s own frontier models.
Engineers aren’t baking a complete, trained model into immutable circuits. Instead, they’re targeting stable computational patterns—attention mechanisms, routing logic, token-generation pipelines—that recur across Gemini’s architecture. By removing the overhead a flexible processor needs to handle any model, Google hopes to slash both the time and energy spent per request.
The chip is not meant to replace TPUs. It would sit alongside them as a specialized inference processor for high-volume Gemini workloads, while TPUs continue handling training, research, and third-party model serving. Google engineers expect to finalize the design over the next year, with deployment targeted for 2028.
What this means for you
Frozen v2 won’t land inside your laptop. Its effects will reach Windows users through the cloud: faster responses from Gemini-powered features in Chrome, Google Workspace, and any service that calls Google’s AI APIs. If the efficiency gains translate into lower operating costs, consumers might see fewer usage caps, shorter wait times, and broader access to advanced features that are currently rationed.
For developers and IT professionals, the biggest near-term impact could be capacity relief. Google Cloud has reportedly been unable to expand some customer deals because its rental computing inventory is tapped out. More efficient inference could free up existing TPUs and GPUs, making it easier for enterprises to secure the compute they need for their own AI projects.
The catch is lock-in. Applications optimized for Gemini’s pricing, latency, and APIs may become expensive to move elsewhere. Organizations that adopt Google’s silicon-to-model stack should plan for portability from day one—using model abstraction layers, open data formats, and fallback providers to avoid being trapped by lower per-token prices.
Why Google is betting on specialized silicon
The move is a direct response to a capacity crisis. Training frontier models grabs headlines, but inference—the act of running a trained model to answer user queries—never stops. Every Google search, Workspace suggestion, and Android interaction that taps Gemini consumes compute. At global scale, the electricity bill and data-center footprint quickly become existential constraints.
Google’s custom silicon journey started with the original TPU in 2015, designed to accelerate the matrix math behind neural networks. Subsequent generations evolved into the Ironwood TPU, a powerful but still general-purpose AI accelerator. Frozen v2 represents a more radical gamble: carving out a chip so tailored to one model family that it behaves almost like a hardware implementation of the software.
If it works, Google could serve ten times more requests from the same power allocation. That doesn’t just lower costs—it creates headroom in data centers that are increasingly limited by local electrical grids, not by physical rack space.
The risks: what could go wrong
Semiconductor design is unforgiving. A chip finalized today must remain relevant in 2028, yet AI model architectures shift fast. If the next major Gemini iteration redesigns its core computation, Frozen v2’s fixed-function blocks could become dead weight. Google must decide now which operations are stable enough to hardwire.
The reported six-to-tenfold improvement also needs scrutiny. “Tokens per watt” is a meaningful metric, but real-world performance depends on context length, batch sizes, latency targets, and system-level overhead. A chip that excels under heavy batching might struggle with interactive responsiveness. Until Google shares benchmarks, the numbers remain a goalpost, not a guarantee.
Manufacturing delays, packaging shortages, and yield issues could push the timeline further out. And even if Frozen v2 succeeds, the rebound effect could erode energy savings: cheaper inference might lead to vastly more inference, keeping total consumption high.
What to do about it today
For most Windows users, the answer is simple: do nothing. Frozen v2 is still a lab project, and any consumer-facing changes are years away.
Enterprise cloud customers, however, should start planning now:
- Monitor Google’s Gemini API pricing and service-level commitments. Any drop in per-token cost or expansion of regional availability could signal that the capacity crunch is easing.
- Evaluate lock-in risk in your AI stack. If you’re building on Gemini, ask how you would switch to an alternative if costs or features shifted unfavorably.
- Watch for hybrid options. Google may eventually offer Frozen-backed capacity as a premium tier for Gemini inference, while keeping current TPU-based services for general-purpose workloads.
Developers who run models on Google Cloud should keep an eye on capacity announcements. If Frozen v2 frees up a significant share of TPU inventory, it might become easier to reserve instances for training and experimentation.
The bigger picture: a hardware-driven AI race
Frozen v2 isn’t happening in isolation. Microsoft is developing its Maia accelerators, Amazon has Trainium and Inferentia, and Nvidia remains the juggernaut with its CUDA ecosystem. Each hyperscaler is trying to break its reliance on external chip suppliers, but Google’s approach is the most audacious: rather than build a better general-purpose accelerator, it’s building a Gemini appliance.
If the gamble pays off, competitors will face pressure to co-design their own models and silicon more tightly. Microsoft, for example, might accelerate work on Maia variants optimized for OpenAI’s models—something that would directly affect the Copilot features shipping with Windows.
For now, the most important date to circle isn’t 2028. It’s the next major Gemini architecture reveal. That will tell us whether Google’s model roadmap is stable enough to make Frozen v2 a breakthrough—or an expensive detour.