NVIDIA Breaks AI Factory Returns Into Output, Hardware Life and Demand

Three pillars of AI factory returns

NVIDIA says AI factory returns are shaped by earning capacity, useful life and demand, and that strength in one factor cannot fully offset weakness in another.

NVIDIA says its AI factories raise earning capacity through higher throughput per megawatt and lower cost per token. SemiAnalysis’ AgentX data shows Vera Rubin NVL72 delivering over 30 times the throughput per megawatt of NVIDIA GB300 NVL72, while cost per million tokens on the DeepSeek V4 Pro model is reduced by up to 45 times.

Epoch AI model shows the effect of hardware life

Epoch AI estimates that a one-gigawatt AI data center requires $38 billion in upfront capital expenditure and $0.9 billion in annual operating expenses. Annualizing capital expenditure over asset lifespans produces a total cost of ownership of $8.5 billion per year. Epoch AI’s model assumes 5 years for IT equipment and 14 years for the facility; shortening IT life to 3 years raises annual cost to $12B, while extending it to 7 years lowers the cost to $7B.

NVIDIA says CUDA runs across generations, while continuing software and kernel optimization can improve existing hardware’s capabilities, so equipment is not stranded when a new architecture arrives. NVIDIA also says its platform supports AI models as well as data processing, scientific computing, simulation and graphics workloads, expanding the demand it can serve.

Dell’s ROI figure carries research qualifications

Dell says more than 4,000 customers have deployed Dell AI Factory with NVIDIA, with early adopters seeing up to 2.6 times ROI in the first year. Dell notes that the first-year ROI figure comes from an Enterprise Strategy Group study commissioned by Dell; the model estimates Llama 3 70B inference and fine-tuning workloads over four years, using an XE9680 server with eight H100 GPUs, and actual results may vary.