The AI Supply Chain's Next Bottleneck Is Not a Chip

The AI buildout looks like a chip story. It is a constraint ladder, and the next rung is not technical.

Three years ago the binding constraint was GPU supply. Hype and capital fixed that. Today, memory is the choke point: high-bandwidth memory, the kind that sits next to a GPU and feeds it without stalling, will not reach supply-demand equilibrium until 2028 or 2029, on Dan Ives' read. SK Hynix is sold out. Micron HBM is locked through 2026. The order book is the evidence. The constraint has moved.

The rung after that is political, not technical. State-level permitting is now a faster-moving variable than factory capacity. Ives called New York's proposed data center moratorium "dangerous" on Anthony Pompliano's show, and the diagnosis travels: the next binding constraint is decided in statehouses and utility commissions, not cleanrooms. Local opposition moves faster than construction.

The rung after that is structural and slow. US grid capacity sits at roughly a quarter of what AI at 20 to 25 percent adoption would need, per Ives' reading of the gap. Three to four times today's build, on a five-to-ten-year clock.

A watchlist, then, not a buy list. HBM order books until 2028. Statehouse permitting fights. Substation queue depth. The AI trade is not a chip trade. It is a sequence of binding constraints, and the next one is decided in Albany.