The Rising Costs and Future Outlook of AI Model Training

TapTechNews July 8th news, according to Tom'sHardware's report today, Dario Amodei, the CEO of AI startup Anthropic, recently said in an interview on a podcast program that currently, the training cost of a model like GPT-4o is about 100 million US dollars, while the training cost of the currently developing AI large model may be as high as 1 billion US dollars (TapTechNews note: about 7.294 billion Chinese yuan currently).

Of course, 1 billion US dollars is not the end point. Dario Amodei predicted that within the next three years, the training cost of AI large models will rise to 10 billion US dollars or even 100 billion US dollars (about 729.428 billion Chinese yuan currently).

Dario Amodei also talked about his view on Artificial General Intelligence (AGI): AGI will not be achieved suddenly, but a gradual development process. New models are built on the development of past models, just like human children continuously learn new knowledge.

The report said that if the computing power of AI models increases by 10 times every year, then it can also be inferred that the hardware needed to train the model must also be at least 10 times stronger. Therefore, hardware is likely to be the biggest cost-driving factor in training AI models. As early as 2023, ChatGPT required more than 30,000 GPUs, and the OpenAI CEO Altman also once confirmed that the training cost of GPT-4 is 100 million US dollars.

At the same time, Elon Musk has also expressed the desire to buy 300,000 NVIDIA B200 AI chips, and OpenAI and Microsoft are also planning to build a 100 billion US dollar AI data center.