According to Tom's Hardware today, AI Startups Anthropic Dario Amodei, CEO of , said in a recent podcast interview that currently, algorithms like GPT-4oModel TrainingcostApproximately US$100 million, and the training cost of large AI models currently under development isCould be as much as $1 billion(Currently approximately RMB 7.294 billion).
Of course, $1 billion is not the end. Dario Amodei predicts that within the next three years,AI Big ModelThe training cost will rise to 10 billion dollars or even 100 billion U.S. dollars (currently about 729.428 billion yuan).
Dario Amodei also talked aboutGeneral artificial intelligence (AGI）AGI “will not happen suddenly”, but will happen over time.Gradual development processNew models build on the development of past models, just as human children continually learn new knowledge.
According to the report, if the computing power of AI models increases 10 times each year, then it can be inferred that the computing power required to train the models will increase by 10 times.The hardware is also at least 10 times more powerfulTherefore, hardware is likely to be the key to training AI models.Biggest cost driver. As early as 2023, ChatGPT will require more than 30,000 GPUs, and OpenAI CEO Altman has confirmed that the cost of training GPT-4 is $100 million.
At the same time, Musk has also expressed his desire to purchase 300,000 Nvidia B200 AI chips, and OpenAI and Microsoft are also planning to build a $100 billion AI data center.