#DeepSeek冲击全球算力 You guys are really making a fuss, the entire Binance Square is filled with similar press releases: from a 90% reduction in training costs, we've won again, the US stock market is doomed, and it even concludes that the cryptocurrency circle is also collapsing, which is ridiculous beyond belief. Today, let's objectively critique Deepseek:
First, regarding Deepseek itself, Deepseek explores the possibility of optimizing training frameworks and model designs under the extreme shortage of computing power to achieve expected tasks. Most importantly, it open-sources the final results back to the community, which is much better than many closed-source companies that create technical barriers. It also debunks the price perceptions created by companies like OpenAI (gpt-o1 requires a $20/month membership to use, while the pro version costs $200). Due to insufficient model data and the loss of some precision from extremely optimized training algorithms, Deepseek's overall user experience still has some shortcomings, but its inference capabilities are already on par with o1.
Next, let's talk about financial issues. In the short term, it will definitely be bad for NVIDIA, as AI companies will pause their aggressive purchasing strategies for computing power and instead return to model optimization. Investors will also reevaluate the return on investment for AI. However, in the long term, it remains positive. If a company develops a cheaper training card, then the negative impact can be understood. But saying this is also negative shows a complete lack of logic, only caring about winning. Here are two short stories:
1. The invention of a more efficient and cheaper steam engine did not lead to a decline in coal sales; rather, coal sales improved because, with the efficiency and price improvements of the steam engine, more small businesses and individual workshops began using steam engines, which increased coal sales.
2. During the prosperous period of British handloom textiles, textile workers thought, wouldn't it be great to invent a weaving machine? This way, a job that originally took 10 hours a day could be done in 3 hours, leaving the remaining time for rest. Later, the weaving machine was indeed invented, and guess what? They still had to operate the machines for 10 hours a day; they just produced more fabric.
The AI industry is similar; with the optimization of training frameworks and models, the entire industry's landscape will change. Some companies that need AI will be fully capable of building their own internal AI training clusters, and even individuals will be able to run localized AI, so in the long run, it still benefits NVIDIA.