Conference Presentation, Panel, Fireside Chat
Eric Schmidt & Andrew Feldman: The Race to SuperIntelligence
- The outlook anticipates a massive global acceleration of AI utility, with inference workloads expected to drive a continued explosion in demand proportional to user count, usage frequency, and compute intensity per interaction.
- Economic and societal benefits are projected to scale globally with "enormous improvements," while the integration of AI into daily life is expected to occur so gradually that it may go unnoticed until it is fully ubiquitous.
- Cost per token served is predicted to decrease by a factor of ten annually, with performance driven by linear algebra acceleration regardless of model architecture, though traditional SaaS metrics are viewed as obsolete in favor of speed-based metrics.
- Scaling laws for deep learning, reinforcement learning, and test-time compute are currently described as being in the steepest part of an S-curve, with the industry in an explosion phase before consolidation, yet the specific point where growth asymptotes remains unknown.
- Future growth will be contingent on high capital allocation and subsidized or inexpensive electricity, with large data centers requiring power equivalent to two large nuclear plants (approx. 3 gigawatts total), creating a significant dependency on nuclear energy deployment.
- The United States is projected to fall significantly behind in energy infrastructure, requiring approximately 92 new nuclear power plants over the next decade to meet demand, having effectively deployed zero large-scale plants in the previous 20 years.
- Europe's ability to build custom infrastructure and silicon is conditional on prioritizing the development of a sustained pool of entrepreneurs and scientists, while alignment challenges regarding self-harmful behavior in self-improving systems remain unsolved and require advanced scientific testing.
- Market risks include severe penalties for latency in AI inference, as customers will immediately switch providers if service is slow, alongside the potential for new AI developments to generate harmful capabilities like cyber or biological attacks.
- Forecasting growth remains "unbelievably challenging" due to unprecedented rates of expansion, necessitating the creation of new metrics to track performance in domains where companies hold a competitive advantage.
- France is identified as possessing some of the cheapest and cleanest energy, specifically nuclear, though heavy taxation currently hinders its utilization, while the industry is characterized as being at the beginning of simultaneous new hardware, software, and fabrication revolutions.