In a recent interview, Anthropic CEO Dario Amodei believes that the current cost of training artificial intelligence models will continue to increase, and even the training cost will increase to $3 billion or even exceed $100 trillion in the next three years.
Dario Amodei said that hardware will become a key factor in the future cost of artificial intelligence training. Taking NVIDIA's "Blackwell" display architecture B200 GPU announced this year as an example, the price is around US$3 to US$4. Tesla CEO Elon Musk plans to spend 30 sets of B200 GPUs to build a large-scale data center to train artificial intelligence models. Microsoft also plans to build a new artificial intelligence data center with OpenAI, with an estimated investment of over one trillion US dollars.
In addition to hardware investment costs, AI model training and subsequent operation also involve energy consumption costs, such as electricity. Therefore, power consumption will become a major challenge for AI development in the future. Other factors, such as how to obtain training data and the associated licensing costs behind that data, could exponentially increase AI training costs.
Dario Amodei's view on the market-discussed artificial general intelligence (AGI) is that such technology will not appear suddenly, but will be developed step by step, just as humans continue to learn as they grow and develop "wisdom" that can be applied.
Prior to this, NVIDIA CEO Jensen Huang also pointed out that the development of artificial intelligence technology will inevitably face the problem of power consumption in the underlying operations, but he expected that it would be able to be improved one by one according to the technological development at that time. Therefore, he believed that as artificial intelligence technology develops, the energy usage problem will be continuously solved.
