Google Develops Frozen V2 Chip to Serve Gemini Model
According to The Information, Google is developing the "Frozen V2" chip to more efficiently serve its Gemini artificial intelligence model.
The new Frozen chip is expected to be 6 to 10 times more efficient than the current TPU, with plans for deployment as early as 2028.
AI computing power capital continues to shift towards self-developed chips, with Google, as a major model operator, becoming a buyer of efficient hardware, putting pressure on external GPU suppliers. The efficiency improvement of Frozen V2 will reduce the training and inference costs of Gemini, strengthening Google's long-term competitiveness in AI infrastructure.
Source: Public Information
ABAB AI Insight
Google has previously launched multiple generations of TPU chips to support its own AI model training and inference. The development of Frozen V2 continues its historical path of reducing reliance on external suppliers through vertical integration of hardware. In terms of capital strategy, Google is concentrating R&D resources on a new generation of efficient chips, aiming for a 6-10 times efficiency improvement to serve the Gemini model. The motivation is to control long-term costs of AI infrastructure and optimize energy consumption, forming a hardware-software closed-loop advantage. Similar to Amazon's self-developed Inferentia/Tranium chips or Microsoft's collaboration with OpenAI for custom hardware, Google is currently transitioning from general-purpose TPU to model-specific optimization in AI chips. Structural Judgment Essentially, this is a technological substitution and industrial chain reconstruction: Frozen V2 significantly enhances efficiency, replacing some roles of general-purpose GPUs in Gemini inference, reconstructing the AI computing power supply chain. This mechanism stems from the energy consumption and cost pressures brought about by the expansion of large models, driving hyperscaler capital towards custom chips. ABAB News · Cognitive Laws
- Self-developed chip efficiency doubles, external reliance halves.
- Computing power costs are ultimately determined by hardware vertical integration.
- The larger the model, the stronger the necessity for custom chips.