While training advanced AI models today costs around a hundred million, the future generation will have a price tag of a billion or even more.

AI models under development today will have training costs of about a billion dollars, claims Dario Amodei, CEO of Anthropic. The price of the most advanced models available today is ten times lower. For example, training GPT-4o cost about 100 million US dollars.

Amodei says in a podcast that he sees costs rising even further. Within a few years, building the models will cost tens to hundreds of billions of dollars. In this scenario, the CEO expects the models to outperform humans in many tasks.

Hardware costs money

Hardware is the biggest cost of training models. A single Nvidia B200 chip, used for the most advanced GPU systems, costs more than $30,000 each. OpenAI’s Sam Altman recently revealed that his company will need around 30,000 GPUs. If those are all B200 chips, the 1 billion mark will be very close, and that’s without additional costs like the physical data center framework, cooling, and power.

However, the biggest cost of the AI ​​boom is not necessarily monetary. The huge GPU clusters consume more electricity than is sustainably available. Google, for example, has a goal of reaching net zero by 2030, but instead saw its greenhouse gas emissions increase by 48 percent in five years due to AI.