Microsoft AI CEO Says Frontier Model Training Could Cost $100 Billion

Microsoft AI CEO Mustafa Suleyman said frontier artificial intelligence training runs could eventually cost as much as $100 billion, underscoring the scale of investment required to develop the most advanced models.

He described possible costs as “many, many tens of billions of dollars if not a hundred billion dollars.” The $100 billion estimate is a forecast, not a budget announced for a Microsoft project.

Suleyman said five or six labs, including Microsoft AI, have the resources to assemble gigawatts of computing power. He did not name the other labs or provide a cost breakdown.

In a Sept. 28 interview, Suleyman also said inference costs — the expense of using a trained model to generate responses — had fallen 300-fold over the previous two years.

His remarks describe contrasting cost trends: frontier training may become more expensive while inference costs have fallen. They do not establish whether lower inference costs will offset the expense of training.