Superintelligence Timelines: Modeling the Trajectory to AGI

The Variance in Expert Predictions

Estimates for when artificial general intelligence will be achieved range from aggressive near-term projections (2026–2028 from figures like Demis Hassabis and Dario Amodei) to skeptical decades-long projections (2040+ from computer scientists citing missing foundational cognitive paradigms).

The Three Physical Scaling Bottlenecks

• Electrical Power & Grid Transmission: Frontier clusters training next-generation systems require gigawatt-scale power installations, necessitating direct co-location with nuclear or dedicated renewable generation assets.
• High-Bandwidth Memory (HBM): Training and running inference on multi-trillion parameter architectures strains global packaging capacity for advanced HBM3e/HBM4 silicon.
• Synthetic Data & Over-Fitting: Having consumed virtually all high-quality human linguistic corpora, frontier labs increasingly rely on synthetic data, risking model collapse without rigorous automated filtering.