Compute (AI Capabilities)

Compute refers to the hardware resources required to train and run AI systems—GPUs, TPUs, and specialized accelerators. Training frontier models costs tens to hundreds of millions of dollars in compute alone.

Compute is uniquely tractable for governance because it is measurable (FLOPs, GPU-hours), concentrated (few chokepoints like ASML, TSMC, NVIDIA), and physical (can be tracked and controlled).

Metric	Score	Notes
Changeability	30	Requires international coordination
X-risk Impact	70	Directly affects capability timelines
Trajectory Impact	80	Primary driver of AI advancement speed
Uncertainty	35	Hardware trends relatively predictable

Related Content

Governance Approaches:

• Compute Governance Overview
• Export Controls
• Compute Monitoring
• Hardware-Enabled Governance

Key Debates:

• Can compute controls effectively slow dangerous AI development?
• Will efficiency gains outpace hardware restrictions?

---

Ratings

Metric	Score	Interpretation
Changeability	30/100	Hard to prevent or redirect
X-risk Impact	70/100	Substantial extinction risk
Trajectory Impact	80/100	Major effect on long-term welfare
Uncertainty	35/100	Moderate uncertainty in estimates