Closing Keynote | Peter Hoose from Meta

The current AI infrastructure buildout represents a historic investment, with spending projected to reach $1 trillion by 2027, far exceeding the inflation-adjusted costs of the Apollo moon missions. This massive shift in capital and engineering focus is transitioning software development from human-authored code to agentic-driven processes, significantly boosting developer productivity. However, this scale introduces critical reliability challenges, as training jobs on 100,000+ GPUs face frequent interruptions that require advanced fault tolerance and diagnostic systems. While AI-driven changes have improved code safety, the sheer volume of updates and the complexity of managing heterogeneous hardware create new operational burdens. Currently in a "Sputnik moment," the industry is only at the beginning of a multi-decade transformation that demands deeper integration between silicon, software, and infrastructure to sustain progress and deliver meaningful real-world applications.