Running a large AI model at scale has become less a scientific feat than a brutal math problem about electricity bills and datacenter rent. Every token Gemini generates costs Google something real—silicon cycles, memory bandwidth, and watts drawn from the wall. As query volume grows, fractions of a cent add up to sums that can swallow margins whole. That quiet urgency is behind Frozen v2, an internal server chip project now taking shape inside Google. Rather than refining its general-purpose Tensor Processing Units for another generation, the company is attempting something far more radical: casting the skeleton of the Gemini model directly into the silicon itself.

From Flexible Accelerators to Model-Specific Silicon

Google’s TPUs have been the workhorses of its infrastructure for nearly a decade. They train models, power search ranking algorithms, and even rent out by the hour to cloud customers including Meta and others looking for an alternative to Nvidia’s GPUs. That versatility is exactly what makes a TPU a TPU. It speaks a general vocabulary of matrix multiplication and memory movement, usable by almost any neural network you can describe in software.

Frozen v2 deliberately trades that flexibility away. The chip is being designed as a domain-specific accelerator whose circuits physically mirror portions of Gemini’s own architecture. Where a TPU fetches instructions and interprets them as software operations, Frozen v2 would burn the model’s structural blueprint—the arrangement of its layers and data paths—directly into the chip’s layout. Google expects this tight marriage of model and metal to make the chip six to ten times more efficient than current TPUs at serving AI responses. Fewer compute steps per query means less time waiting for a token to appear, and far less energy spent generating it.

This is not merely a faster version of the same idea. It is a different category of chip, one that trades generality for devotion to a single model family.

Why the First “Frozen” Melted

This approach has roots in an earlier concept attributed to Jeff Dean, Chief Scientist at Google DeepMind. The original “Frozen” proposal suggested pushing specialization even further by hardcoding not just the architecture, but the actual model weights—the billions of tuned parameters that constitute Gemini’s learned behavior—directly into the chip itself.

The logic was sound. If you know exactly which numbers the model will use, why bother fetching them from external memory? You could etch them into the transistors and eliminate entire categories of delay.

The problem was permanence. AI models do not stand still. Google updates Gemini continuously, retraining on new data, adjusting parameters, releasing improved versions. A chip with weights frozen in silicon would become a paperweight the moment a new model revision shipped. That lack of flexibility killed the original concept.

Architecture Without the Anchor

Frozen v2 solves the obsolescence trap by hardcoding the architecture while leaving the weights free to change. Think of it as pouring a custom racetrack instead of welding the car to the road. The shape of the circuit stays fixed, optimized for Gemini’s specific patterns of computation, but the contents flowing through those circuits can be refreshed by loading new weights from memory.

This distinction matters in practice. When engineers train a new Gemini checkpoint, they can deploy it to Frozen v2 hardware without fabricating a new chip. The exact degree of hardcoding is still an open question inside Google; teams must decide precisely which structural elements deserve silicon immortality and which should remain configurable. But the principle is settled. By freezing the shape and fluidly swapping the parameters, Google keeps the efficiency upside without sacrificing the ability to iterate.

The Economics of Keeping It In-House

There is another reason you will not see Frozen v2 listed on Google Cloud’s pricing sheet. Because the chip is so intimately shaped around Gemini’s internals, it would make little sense to outside developers running PyTorch or custom Transformer variants. Google has no plans to sell it as a general-purpose product. It will remain an internal tool, aimed squarely at the crushing demand for inference capacity inside Google’s own datacenters.

这一选择反映了一个残酷的经济现实。在当前的生成式 AI 市场中，模型能力正在迅速趋同。竞争对手之间的差距往往取决于谁能以最低的单 token 成本运行最大的模型。推理不再是训练之后的次要考虑；对于像 Gemini 这样被广泛使用的产品来说，推理是主要的支出。如果 Frozen v2 能将该项支出降低六倍或更多，Google 将获得竞争对手难以轻易企及的优势。它既可以将节省下来的成本转化为利润，也可以通过降低 API 消费者和产品集成的价格来回馈用户，从而向 OpenAI、Anthropic 等对手施压。

这向行业传递了什么信号

Google 的这一举动也暗示了更广泛的硬件战略走向。多年来，标准的做法是构建尽可能灵活的加速器，并让软件来处理专业化需求。Nvidia 的 GPU 占据主导地位，因为它们可以运行从分子动力学到视频游戏再到大语言模型的一切内容。Google 自己的 TPU 最初也是秉承这种广泛适用性的理念设计的。

Frozen v2 打破了这一传统。这是一种承认：当单一模型家族驱动了足够的查询量时，为该模型量身定制的定制化芯片可以带来数倍的回报。其他超大规模云服务商也遵循类似的逻辑——例如 Amazon 的 Trainium 和 Inferentia 芯片——但 Google 的方法更进一步，它是围绕特定的模型架构而非通用的网络类别来协同设计硬件。

当然，风险在于僵化。如果 Gemini 的架构向硬编码电路无法适应的方向演进，Google 可能会发现自己拥有的昂贵芯片无法运行其最新的构想。这正是为什么“仅针对架构的折中方案”至关重要的原因。它提供了一条中间道路：既有足够的专业化程度来获取显著的效率提升，又有足够的灵活性来避免让公司陷入困境。

真正的启示

理解 Frozen v2 的最佳方式不是将其视为一个芯片发布，而是将其视为对 AI 竞争未来格局的一次战略押注。Google 正在赌，最终的赢家不仅会构建出最好的模型，还将拥有整个技术栈——从模型的蓝图一直到穿过晶体管的电子。如果该项目取得成功，其回报不会体现在基准测试分数中。它会体现在季度财报的成本栏中，每百万 token 节省下来的几美分，就足以重新划定生成式 AI 商业可行性的边界。