NVIDIA's blog post lays out a framework for AI factory economics, arguing that returns depend on three factors: earning capacity, useful life, and demand. The company claims its systems are engineered to maximize all three simultaneously, and that weakness in one cannot be offset by strength in another.
On productivity, NVIDIA cites SemiAnalysis data showing its next-generation Vera Rubin NVL72 delivers over 30x higher throughput per megawatt and up to 45x lower cost per million tokens compared to the GB300 NVL72 on a specific model. The post argues that cheaper tokens expand demand rather than shrink it, as more use cases become economical.
Durability is supported by examples like the A100, still in commercial service six years after launch, and CoreWeave extending bookings for 2020-era units through 2029. The post also notes that depreciation schedules keep extending, and that CUDA compatibility prevents stranded hardware when new architectures arrive.
Fungibility means the same infrastructure runs every type of AI and non-AI workload, from data processing to scientific computing. NVIDIA claims this broadens demand and extends useful life, making its GPUs general-purpose accelerated computing rather than single-purpose ASICs. This is a single-source analysis based solely on NVIDIA's claims; no independent verification is provided.