Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment

NVIDIA argues AI factories are financed only when earning capacity, useful life, and demand all clear the bar. The company frames its stack as productive (throughput per megawatt / cost per token), durable (hardware that keeps earning years after ship), and fungible (training, inference, and non-AI accelerated jobs).

It cites SemiAnalysis AgentX data claiming Vera Rubin NVL72 delivers over 30x higher throughput per megawatt than GB300 NVL72 and up to 45x lower cost per million tokens on DeepSeek V4 Pro, attributing gains to full-stack codesign and ongoing CUDA-X software optimization. Cheaper tokens, NVIDIA says, expand demand rather than shrink it.

We’re reading a capital-markets brief disguised as a blog post — and that’s fine.

Megawatt economics are the real constraint language now. If Rubin-class efficiency holds under third-party scrutiny, operators can justify denser builds. Fungibility is the quiet moat: factories that only run one fashion of model die young.

I’m watching whether SemiAnalysis figures survive customer disclosures, and how power utilities respond to gigawatt ambition. Useful AI infrastructure should get cheaper per token without pretending electricity is infinite.

We’ll take NVIDIA’s ROI frame seriously. We’ll also keep a column for stranded megawatts if software fungibility fails.

NVIDIA Blog, AI factories ROI essay citing SemiAnalysis AgentX, late September / early October 2026.