After Google launched the new generation of AI Gemini 3, the stock price rose day after day, from $285 to 325 yuan. The reason is not only that Gemini 3's ability to generate pictures and small games is epoch-making, but also its TPU threatens Nvidia's exclusive status of "AI selling shovels" may quietly dig another channel in Nvidia's moat.
Selling shovels under the tide of gold is the most stable business. Nvidia has always built an AI moat for shovels with its powerful GPU and deep CUDA technology. CEO Huang once boasted that even if competitors sell chips for free, for AI companies (technical costs) "It's also too expensive." However, since Google launched Gemini 3, Wall Street has been more cautious about Nvidia, and the stock price has been under pressure and repeatedly down.
the training and reasoning of the Gemini 3 Pro released by Google this time are completely based on the self-developed TPU cluster, which represents the trillion-based parameter-level MoE (mixed expert model) and the ultra-long context (Infinite Context ) Scene, the ASIC chip scheme designed by software and hardware has been successfully run. At present, the mainstream large model architecture tends to be standardized. Llama, Gemini, etc. are all based on Transformer architecture, and the core of Transformer is the large-scale matrix multiplication plus Attention mechanism, which is what TPU pulsating arrays are best at. The calculation mode of. This means that the training of top large models is no longer indispensable for Nvidia GPU.
In addition, in terms of training consumption, Google TPU v6e has approached Nvidia's Blackwell in terms of unit computing power capital investment (TOPS/USD). According to the current public data and evaluation, under the sub-item of "Training Standard Transformer / LLM Large Model", the performance-consumption ratio of TPU has caught up with the same generation of data center GPU, and even more advantages.
At the same time, Google TPU has begun to commercialize, not only opening TPU v7 Ironwood in the Cloud, but also reaching a million-card-level TPU cooperation agreement with Anthropic, all of which have brought a 10 billion-dollar-level camp after putting into production. Receive. Canada posted that Meta is negotiating with Google about the use of TPU chips worth billions of dollars. All of them threaten the 90% market share of Nvidia's monopoly level.
The potential of $谷歌(GOOG)$ is unlimited
Modify on 2025-12-04 09:56
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
