Progress in AI is usually reported as a sequence of research results. The rate at which those results arrive is governed by how much specialised computing hardware exists, and that quantity is constrained at several physical points.
Training is a hardware-bound activity
Training a frontier model means running an enormous number of matrix operations, and the wall-clock time depends on how many accelerators can be applied at once.
A lab with access to a larger cluster can run more experiments in a year, and experiments are how the useful configurations are discovered.
Research productivity therefore tracks hardware allocation closely, which is why capability concentrates among organisations that secured supply early.
Advanced packaging is the narrow point
The constraint is generally not the logic wafers themselves but the step that bonds high-bandwidth memory to the processor.
That packaging capacity is concentrated in very few facilities, and building more takes years because the equipment involved has its own long lead times.
Memory supply is similarly concentrated among a small number of manufacturers, and expanding it requires committing capital well before demand is confirmed.
Power arrives more slowly than hardware
Even where accelerators are available, they must be installed somewhere with sufficient electrical supply and cooling.
Grid connections at the scale required are negotiated with utilities and take years, and the transmission equipment involved is itself in short supply.
This is why announcements about new capacity are increasingly about substations and interconnection agreements rather than about chips.
Export controls fragment the market
Restrictions on where the most capable accelerators may be sold have created separate regional markets with different available hardware.
Organisations under restriction adapt by training smaller models more efficiently, and by pursuing architectural work that reduces hardware demand.
Constraint has produced real efficiency research, which is a genuine effect rather than a rhetorical one, though it has not closed the gap in absolute capability.
Inference is quietly taking the larger share
Attention focuses on training runs, and the sustained demand comes from serving models to users, which continues indefinitely once a model exists.
As deployment has grown, a rising share of installed hardware serves requests rather than training anything new.
That shift changes what is scarce: training capacity determines what becomes possible, while serving capacity determines how much of it anyone can actually use.