Nvidia published metrics detailing how AI factory operators shape return on investment through earning capacity, useful life, and demand. Each megawatt factory costs roughly $60 million, according to Nvidia data.
Nvidia stated that power is the binding constraint on an AI factory. SemiAnalysis AgentX data shows Nvidia Vera Rubin NVL72 systems deliver over 30 times higher throughput per megawatt than Nvidia GB300 NVL72. The systems also deliver up to 45 times lower cost per million tokens on the DeepSeek V4 Pro model.
Hardware lifespan
Nvidia hardware keeps earning years after shipping. The Nvidia A100 graphics processing unit shipped in 2020 and remains in commercial service six years later. CoreWeave extended bookings for units introduced in 2020 through 2029.
Barkr estimates useful life at five to six years for an eight-graphics processing unit H100 system. Barkr estimates nine to 10 years for GB300 NVL72 systems based on resale values. Silicon Data reported that a six-year-old A100 unit is worth a quarter of its original cost.
Ornn Data found the market paying 80% as much to rent an A100 unit on a five-year contract as on a one-month contract. Major operators have extended depreciation schedules for their servers over time.
Customer deployments
Lilly runs protein, small-molecule, and genomics models on a 1,016-graphics processing unit on-premises cluster. Pinterest deploys a vision language model across 14,000 units spanning Blackwell, Hopper, and earlier architectures.
Revolut processes transaction records using cuDF and trains foundation models on an AI cloud. Runway trains a world model on Nvidia Hopper and serves it on the Blackwell platform.
Texas A&M University operates molecular simulation and drug discovery on a supercomputer at 95% to 98% utilization across 26 projects and seven institutions.
Nvidia founder and CEO Jensen Huang will discuss AI factories during his GTC Berlin keynote on Wednesday, October 21, 2026, at 11 a.m. CEST.
