Although the OpenAI CEO denied it before, the industry still believes that they are already training the GPT-5 large model, which will be more than 10 times the size of GPT-4, but this also means more money, especially the graphics card used for training AI is extremely scarce. There are so many AI large-scale models in the world, how many AI graphics cards are used is the secret of each company, and accurate data is rarely disclosed. GPT-4 is guessed to be trained on 10,000-25,000 A100 graphics cards. The graphics card required for GPT-5 Still a fan, Elon Musk estimates 30,000-50,000 H100 graphics cards, which have much stronger performance.
In addition to OpenAI, the demand for high-performance AI graphics cards from other companies is also high. Meta, Google, Microsoft, Apple, Tesla and Elon Musk are all engaged in various AI, and the total demand is predicted to be as high as 432,000 H100 graphics cards. , worth more than $15 billion.
This is just the demand of the American technology industry analyzed by foreign netizens. In fact, the domestic demand is not less than that of the United States. Even if you can only buy the special version of the A800 and H800 accelerator cards, this does not prevent domestic companies from investing in large AI models and snapping up AI at a higher price. graphics card.
In the current market, only NVIDIA can meet the needs of AI graphics cards. AMD and Intel’s AI graphics cards not only have performance and ecological problems, but also cannot keep up with the supply. AMD’s killer MI300X will not be shipped until the end of the year. It will be launched in large quantities in 2024.
In the final analysis, this wave is another big profit for NVIDIA. The demand for AI graphics cards will continue until the end of 2024, and the supply will be in short supply in this year and a half. The price of H100 graphics cards starts at 250,000 CNY.
