What's this about?
Arvind Narayanan worked through what data-centre bans or moratoria actually do to AI progress. His method: because inference efficiency at a given capability level improves steadily and will likely keep doing so, blocking a given amount of capacity translates into a modest delay rather than a stop. Mollick reached the same conclusion independently.
What exactly does "1 GW of capacity blocked" mean?
Data centres are measured in power, not machines, because electricity is the binding constraint. A gigawatt is roughly a large power station's output. Narayanan's move is to convert that unintuitive figure into time: if efficiency at a fixed capability improves at a steady rate, blocked capacity is equivalent to setting the clock back by however long it takes efficiency to make up the difference.
What's the community saying?
157 likes on the analysis, 117 on the concurrence
- The conclusion both reached: state-level bans have very little effect on overall AI progress.
- The important corollary, stated explicitly — if your concern is AI development, concentration of power, or future use rather than "is there a data centre near me", this is not a substitute for policy.
- The methodological caveat is that it rests on efficiency trends continuing, which the Jalapeño chip item above supports and which could also plateau.