NVIDIA outlines three drivers of AI factory returns: productivity, durability and flexibility

NVIDIA has presented a framework for evaluating AI factory returns around three attributes: productivity, durability and fungibility. The company says a one-megawatt facility costs about $60 million, making a clear return on investment essential for operators.

Returns depend on how many tokens a system can produce, how long its hardware remains economically useful and how much demand exists for its capacity. Strength in one area cannot fully compensate for weakness in another.

Vera Rubin lowers token costs

SemiAnalysis AgentX data cited by NVIDIA shows Vera Rubin NVL72 systems delivering more than 30 times the throughput per megawatt of GB300 NVL72 systems. On the DeepSeek V4 Pro model, the company says the cost per million tokens can be up to 45 times lower.

NVIDIA attributes the gains to joint optimization across models, software, compute, networking and memory. It also argues that cheaper tokens can make more use cases economical, increasing total demand for compute.

Older systems keep earning

NVIDIA points to the A100, launched in 2020 and still in commercial use. CoreWeave has extended bookings for units introduced that year through 2029. Barkr estimates useful life at five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72.

Silicon Data values a six-year-old A100 at one quarter of its original price. Ornn Data says a five-year A100 rental contract commands 80% of the monthly rate. NVIDIA says CUDA compatibility across generations and ongoing software optimization allow existing systems to remain productive after newer architectures arrive.

One platform for varied workloads

NVIDIA says its infrastructure supports machine learning, generative and physical AI, data processing, scientific computing, simulation and graphics. CUDA-X includes more than 1,000 libraries and models, while more than 10 million developers build on the platform.

The company cited Lilly running protein, small-molecule and genomics models on a 1,016-GPU cluster; Pinterest using 14,000 GPUs across several generations; and Texas A&M University reaching 95-98% utilization across 26 projects and seven institutions.

NVIDIA also described non-AI uses: Cosm applies its infrastructure to video and graphics, Dassault Systèmes to digital-twin simulations, and Unilever to product imagery. NVIDIA said CEO Jensen Huang’s GTC Berlin keynote is scheduled for October 21 at 11 a.m. CEST (9 a.m. UTC).