Capital move to bottelneck

If Inference Economics Drive AI Margins, Capital Will Follow Bandwidth.

If cost per token determines AI profitability, then capital allocation shifts away from raw compute and toward:

• High-bandwidth memory (HBM) supply
• Rack-scale interconnect architectures
• Optical networking & switching
• Power-dense data center design
• Inference optimization software

Inference is increasingly memory-bound and network-bound

That means future AI capex won’t just be “more GPUs.”
It will be:

More bandwidth per rack.
More efficient distributed stacks.
More power and cooling per square foot.

The capital pattern is now clear

Inference utilization determines value capture

architecture determines inference economics

the real question for investors isn’t:

Who has the biggest model?

It’s:

Who reduces cost per token the fastest — and which infrastructure layers benefit from that shift?

Training built the engine.
Inference is allocating the capital.