Compute Leverage

Fresh3d

Who controls the compute a frontier training run needs — read at two grains. Across the eleven layers of the stack, whether a single actor can veto a run and whether anything checks that power; across the facilities and countries that host it, how much capacity sits under whose jurisdiction and where allied export control actually binds. Leverage is switching cost times redeployment time throughout, never market share.

The brief

One screen, plain English: how many chokepoints can each halt the next frontier run, which of them answers to no one, and which stay locked past five years — the finding without the axis machinery.

The finding

Eight chokepoints can each halt the world’s next frontier AI. One answers to no one.

Training a frontier model runs through a long supply chain: the chips, the handful of machines that can make them, the packaging that fuses them, and the power to run it all. At eight points along that chain, a single player can stall the work for a year or more. Seven of the eight sit under some check, whether an export-control rule or a public body. The last answers to no one.

That one is advanced packaging: the step that fuses bare silicon into a working accelerator, controlled by TSMC. No export rule reaches it and no antitrust case touches it; it answers only to the customers in its queue.

Chip lithographythe machines that pattern leading-edge chipsChecked by export-control law
Leading-edge fabricationthe fabs that print the smallest transistorsChecked by export-control law
Chip-design softwarethe EDA toolchains every leading-edge chip is designed onChecked by export-control law
High-bandwidth memorythe stacked memory every accelerator needsChecked by export-control law
Advanced packagingthe step that fuses bare silicon into a working acceleratorChecked by no one
AI acceleratorsthe chips that do the trainingChecked by export-control law
In-server interconnectthe high-speed links between chips in one boxChecked by export-control law
Grid powergetting hundreds of megawatts connectedChecked by a public body

Most of these loosen within five years as rivals qualify. Two do not: chip lithography and chip-design software, where no second supplier is in sight.

This is the brief. Open the Leverage view for the full instrument: every holder, the leverage behind each verdict, the sources, and the methodology.

Built for the AI governance research community

Eight chokepoints can each halt the world’s next frontier AI. One answers to no one.

Chip lithographythe machines that pattern leading-edge chipsChecked by export-control law

Leading-edge fabricationthe fabs that print the smallest transistorsChecked by export-control law

Chip-design softwarethe EDA toolchains every leading-edge chip is designed onChecked by export-control law

High-bandwidth memorythe stacked memory every accelerator needsChecked by export-control law

Advanced packagingthe step that fuses bare silicon into a working acceleratorChecked by no one

AI acceleratorsthe chips that do the trainingChecked by export-control law

In-server interconnectthe high-speed links between chips in one boxChecked by export-control law

Grid powergetting hundreds of megawatts connectedChecked by a public body

Most of these loosen within five years as rivals qualify. Two do not: chip lithography and chip-design software, where no second supplier is in sight.

This is the brief. Open the Leverage view for the full instrument: every holder, the leverage behind each verdict, the sources, and the methodology.