Microsoft AI CEO Predicts $100 Billion Training Costs for Frontier Models

Microsoft AI CEO Mustafa Suleyman expects the price of training frontier AI models to surge to tens of billions of dollars, potentially reaching $100 billion per run in the next two years. Suleyman noted that only five or six laboratories globally possess the capital required to sustain such expenditures.

Speaking on September 28, 2026, Suleyman linked these rising costs to the need for increased compute power to accelerate development. He highlighted that training compute for frontier models has expanded a trillion-fold over the last 15 years. Conversely, he noted that inference costs have dropped 300-fold during the past two years, making the use of AI significantly cheaper.

Suleyman previously predicted in December 2025 that maintaining competitiveness in the AI sector would require hundreds of billions of dollars in investment over a five to ten-year period. He currently characterizes Microsoft’s AI operations as a modern construction company focused on building large-scale compute infrastructure.

The shift toward self-sufficiency in frontier model development follows a restructuring of the partnership between Microsoft and OpenAI. As costs climb, Suleyman anticipates that the number of organizations capable of competing at the highest level will consolidate to a small group of five or six entities.

Alongside the focus on infrastructure spending, Suleyman emphasized the need for increased oversight of AI capabilities. He cited the emergence of autonomous hackers and sophisticated AI agents as a watershed moment that necessitates greater scrutiny of system behaviors.