Insult to injury but those giant AI datacenters are really only needed to train the models, not actually run them.

All they really need is some conventional high bandwidth interconnect GPU servers and some quantization magic to run in a regular old data center.