Mustafa Suleyman, CEO of Microsoft AI, stated that the cost of training a single frontier AI model could reach tens of billions or even $100 billion in the next two years. He estimates that only five or six laboratories globally possess the financial resources to sustain such investments.
Suleyman highlighted that training compute requirements have increased by a trillion-fold over the last 15 years. Conversely, the cost of inference has decreased 300-fold in the past two years, making the deployment of AI models significantly more affordable.
The executive noted that Microsoft is shifting its strategy toward self-sufficiency for frontier models following a restructuring of its partnership with OpenAI. He compared Microsoft AI operations to a construction firm focused on building large-scale compute infrastructure.
Suleyman emphasized the need for increased oversight of AI systems, citing the emergence of autonomous hackers and advanced agents. He argued that the rapid development of these technologies represents a critical moment requiring closer scrutiny of system capabilities.
