Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment
NVIDIA published a blog post stating that AI factories, built by the megawatt or gigawatt, cost roughly $60 million per megawatt and that operators commit capital at that scale only with a clear view of return on investment. It names three factors shaping returns: earning capacity, useful life and demand.
NVIDIA describes its AI factories as productive, durable and fungible. It says productive means the highest throughput per megawatt and lowest cost per token; durable means NVIDIA GPUs and systems keep earning years after they ship; fungible means they run every type of AI, in every phase and place, plus many non-AI workloads.
Citing SemiAnalysis AgentX data, NVIDIA says Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than GB300 NVL72 and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model. The post cites a SemiAnalysis AgentX analysis about GB300 NVL72 performance for GLM 5.3. It attributes the gains to codesign across models, workloads, software, compute, networking and memory.
On durability, NVIDIA says the A100 GPU shipped in 2020 and remains in commercial service six years later, and that CoreWeave recently extended bookings for units introduced in 2020 through 2029. It cites a September 2026 Sprout analysis, "The Productive Life of a Data Center GPU," saying every major operator has extended server life, and that Microsoft's NVIDIA V100 fleet ran 8.4 years against a six-year book life. It also cites Barkr, Silicon Data and Ornn Data figures on useful life, resale value and rental pricing for A100 GPUs.
NVIDIA lists customer examples: Lilly on a 1,016-GPU on-premises cluster; Pinterest post-training and deploying a vision language model across 14,000 GPUs spanning Blackwell, Hopper and earlier architectures; Revolut using cuDF; Runway training on Hopper and serving on Blackwell; and Texas A&M University at 95-98% utilization. It also cites Cosm, Dassault Systèmes and Unilever for non-AI workloads. NVIDIA said CEO Jensen Huang's GTC Berlin keynote is scheduled for Wednesday, Oct. 21, at 11 a.m. CEST.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
AI factories are built by the megawatt, even by the gigawatt. Each megawatt factory costs roughly $60 million, and AI factory operators will only commit capital on that scale with a clear view of the return on investment. Three key things shape AI factory returns: Earning capacity: What the factory could earn in a year […]