NVIDIA on October 1 published a blog post that argues its AI factory architecture maximizes return on investment by combining high tokens-per-megawatt throughput, multi-year hardware earning life, and workload fungibility.
The argument arrives as AI factory operators commit roughly $60 million per megawatt and demand clear payback timelines. Automatica reported in September that NVIDIA’s DSX AI Factory Platform was engineered to treat each megawatt as a critical resource; the new post frames how that architecture drives investor returns.
Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than the earlier GB300 NVL72 and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model, according to NVIDIA, which cites SemiAnalysis AgentX data. The gains come from full-stack codesign—optimizing models, software, compute, networking and memory together—rather than a single component.
On durability, the post states that A100 GPUs shipped in 2020 remain in commercial service, with CoreWeave recently extending bookings through 2029. A September 2026 Sprout analysis tracked how major operators have lengthened depreciation schedules; Barkr pegs useful life at five to six years for eight-GPU H100 systems and nine to ten years for GB300 NVL72, based on resale values. NVIDIA says continuous software optimization keeps installed hardware productive, and CUDA’s cross-generational support means nothing is stranded when a new architecture ships.
For workload versatility, the company points to more than 1,000 ready-made CUDA-X libraries and customer deployments: Lilly runs protein and agentic AI on a 1,016-GPU cluster; Pinterest post-trains vision language models across 14,000 GPUs spanning Blackwell, Hopper and earlier architectures; Revolut uses cuDF for billion-record data processing alongside foundational model work; Runway trains on Hopper and serves on Blackwell; Texas A&M operates molecular simulations and AI drug discovery at 95–98% utilization across 26 projects; and Dassault Systèmes powers virtual twin simulations for aircraft certification and vehicle design.
All claims in the blog are NVIDIA’s own, drawn from vendor-provided data and un-named analyst reports. No independent test results or third-party benchmark verifications are provided. The post does not detail the configuration of the comparison workloads or address how depreciation assumptions might shift if token prices fall faster than the hardware ages.
NVIDIA founder and CEO Jensen Huang is expected to discuss AI factories further during a GTC Berlin keynote on October 21.