dwarkesh

Grindability drives AI gains: cheap parallel math and code rollouts burn ever more GPU compute, the pick-and-shovel layer.

If computer-use progress means grinding thousands of parallel rollouts, the GPU cloud renting that compute compounds.

AI keeps demand for compute insatiable, so the share of the economy spent on compute rises and GPU rents stay high.

A Second Cold War redirects capital from trade efficiency to rearmament, and defense contractors' order books compound.

China's economy runs on sea lanes that close in any conflict, a permanent geographic discount on its assets.

Privatized AI value is now coming public; Anthropic's S-1 starts the re-rating of the biggest private AI name.

If AI compute demand never satiates, GPU pricing power persists and NVIDIA's margins and revenue keep compounding.

If intelligence falls out of raw scale, the compute owner compounds while efficiency tricks get arbitraged away.

RL's bits-per-flop problem pushes spend off the training update onto rollout/inference compute, where CoreWeave bills.

HBM bandwidth is the hard ceiling on AI scaling, and the dominant supplier owns the scarce asset and the pricing.

Rack-scale NVLink fabric (8 to 72 to 576 GPUs) is the real moat; rivals can't drop-in match it, so NVIDIA's premium holds.