Why CPO matters

AI scaling is no longer compute-bound alone. It is interconnect-bound. As clusters move from tens of thousands to millions of GPUs, copper becomes a bottleneck in:

Power consumption

Latency

Signal integrity

CPO reduces power per bit and enables denser rack-scale designs. If NVIDIA controls optical capacity into 2027–2030, it protects the next scaling phase of Blackwell successors.

Is there a “second curve”?

The first curve was training acceleration.

The second curve is likely:

1. Inference at planetary scale

2. Network dominance via NVLink + optical fabric

3. Full-stack integration from silicon to system to interconnect

If NVIDIA owns the fabric layer, it widens the moat beyond GPUs.

Valuation question

A trillion-dollar valuation requires:

Sustained data centre revenue growth

High-margin inference monetisation

No severe hyperscaler insourcing shock

The risk is not technology. It is capex discipline from customers. If AI spend shifts from “grab compute” to “prove ROI”, multiples compress before fundamentals fail.

Long term, I remain structurally constructive.

But the stock’s path will be cyclical, even if the platform thesis compounds.