Training AI costs around $1 billion now, can go up to $100 billion in 3 years: Anthropic CEO

Hardware and infrastructural costs like power drive up training costs for AI models in multiples.

AI models which are in development today can cost up to $1 billion to train, Dario Amodei, the CEO of Anthropic, a US-based artificial intelligence (AI) startup, said in the In Good Company podcast.

Traiing costs for AI is expected to multiply over the years (Unsplash)

What are the current costs and expected future costs to train AI?

Current models like ChatGPT-4o cost about $100 million to train, but this cost for training is expected to reach $10-100 billion in as little as three years, he said.

The development of AI from generative artificial intelligence (like ChatGPT) to artificial general intelligence (AGI) is the reason for the increase in training cost, Tom’s Hardware wrote, quoting Amodei who compared this to how a human child learns.

What are the reasons for cost of training AI to go up in the future?

It was reported back in 2023 that ChatGPT would require more than 30,000 GPUs, with Sam Altman confirming that ChatGPT-4 cost $100 million to train, according to the article, which added that over 3.8 million GPUs were delivered to data centers last year, with Nvidia's latest B200 AI chip costing around $30,000 to $40,000.

Elon Musk is keen on purchasing 300,000 of these chips and Microsoft and OpenAI are reportedly planning to build a $100 billion data center for AI.

The total power required for all the GPUs sold last year could be enough for 1.3 million homes which further adds to the cost, the report read.

All this is despite Google's DeepMind unveiling a new method called joint example selection (JEST) to accelerate AI training with up to 13 times fewer iterations and 10 times less computation, which also reduces the resources and time needed to do the job.