AI/SI infrastructure analysis should extend beyond accelerator demand. A system can have available compute and still face bottlenecks elsewhere in the stack.
Networking and data movement
Large-scale training depends on moving enormous amounts of data efficiently between systems. Networking performance can influence utilization and training economics.
Power and cooling
Access to electricity, grid interconnection, and cooling can constrain how quickly new data-center capacity comes online.
Supply-chain dependencies
Optical components, memory, networking hardware, transformers, and other infrastructure can become critical even when they receive less attention than GPUs.
The investment question is not only who sells compute. It is which bottlenecks become scarce as AI/SI infrastructure scales.