$NVIDIA(NVDA.US)

Post 2 Episode 24

Context: What I find increasingly important about NVIDIA is that the AI conversation is shifting from “Who can train the biggest model?” toward “Who can run intelligence most economically at scale?” NVIDIA’s latest Vera Rubin MLPerf results showed up to 3.7× higher inference throughput than GB300 NVL72, while management says Rubin can deliver roughly 30× higher throughput per megawatt and 35× lower token cost versus Grace Blackwell Ultra for relevant workloads.

My trade: That makes my NVIDIA thesis increasingly about AI economics rather than benchmark supremacy. As inference and agentic workloads become a larger share of AI spending, customers will care about tokens generated per dollar, per watt and per rack. If NVIDIA can keep lowering those costs through co-design across GPUs, CPUs, networking and software, each generation gives customers an economic reason to upgrade — not simply a technological one.

Takeaway: The long-term winner in AI may not simply be the company with the fastest chip. It may be the one that makes intelligence cheapest to produce at massive scale. NVIDIA increasingly looks like it is designing around exactly that objective. However, Nvidia is just a small percentage of my portfolio.

@Captain's Treasure