THE COMPUTE FRONTIER
The accelerator supply chain that decides who can train the largest models.
Summary
Frontier AI is gated by access to advanced accelerators, the power to run them, and the fabs that make them. This File tracks how compute became a strategic resource priced and rationed like one.
The bottleneck is not ideas but the physical stack beneath them: silicon, memory bandwidth, interconnect, and electricity.
Timeline
- 2012
GPU inflection
Deep learning demonstrates that general-purpose GPUs vastly outperform CPUs for training, reorienting the industry.
- 2023
Accelerator scarcity
Demand for leading accelerators outstrips supply; allocation becomes a competitive advantage in itself.
Key Actors
Designers of frontier training hardware
Largest buyers and operators of compute
Constraint on where compute can physically grow
Related Files
Signals
Reported leading-edge fab utilization remains near capacity despite a broad hardware glut elsewhere, suggesting the chokepoint in FILE #001 is tightening, not easing.
Large power transformer lead times are again being quoted in years, reinforcing the binding constraint identified in FILE #003.
Accelerator allocation is increasingly negotiated as multi-year supply commitments rather than spot purchases — compute behaving like a strategic reserve (FILE #002).
Datacenter siting decisions are increasingly driven by grid interconnection availability rather than land or tax — FILE #002 and FILE #003 are converging.
Open Questions
- ?Does power, not silicon, become the binding constraint on compute within the decade?
Sources
- Compute and capability scaling — public research · 2023
