So energy efficiency is literally the bottleneck of the entire AI industry right now. That makes sense. But building a fast, efficient chip doesn't guarantee a monopoly. AMD makes fast chips. Let's talk about the moat. The report points to the CUDA software ecosystem and their networking hardware as the real reason nobody can catch them. Hardware is only as good as the software that can run on it.

Broadcom is the defensive derivative play here. They provide the IP and design services to help Google or Meta build those custom chips. However, the caveat is that captive hyperscaler silicon lacks the flexibility external developers need. And it currently cannot match the raw compute power of NVIDIA's flagship GPUs for trading the very largest frontier models. But for running inference on established models, it is a formidable threat.

That is an excellent analogy. The switching costs are prohibitive. If an enterprise has spent millions of dollars and thousands of hours building their AI infrastructure on CDA, moving to a competitor's hardware means rewriting foundational code bases. Yeah, that's brutal. It's too risky, too expensive, and too slow. And then NVIDIA compounds that software moat with their networking hardware. NVLink and SpectreMax Ethernet. Right. They realize that when you're linking tens of thousands of GPUs together for an AI factory, the bottleneck isn't the chip itself. It's how fast the chips can talk to each other. By dominating the interconnect technology, they aren't just selling individual chips anymore.