AI Infrastructure Is the New Bottleneck

Everyone watching AI right now is watching the models, how fast they’re improving, and which lab is ahead. That’s where most of the progress has been, but for years, advancement was limited by model architecture and by lack of data and compute. Now, the constraint is moving again.

Data centers can’t be built out fast enough. Power is part of it, but so are permits, equipment, and basic construction realities. The grid is slow to connect new loads, and in some places it takes years. Key electrical components have long lead times. On the hardware side, supply is still tight. Chips are still hard to get, and the memory these systems depend on is largely spoken for.

The models will keep getting better. But model quality isn’t what limits how much AI actually gets deployed. Deployment runs on permits, and concrete and electrical equipment that moves nothing like software does.

The last big challenge in AI was data and computing power. Now, the bottleneck is in the real world, and you can’t just scale it up whenever you want.

Originally published on Substack.