AI Briefing
KOSign in

NVIDIA Details How Productivity, Durability, and Fungibility Maximize AI Factory ROI

·2026.10.01 22:00

Key point

SemiAnalysis data shows NVIDIA Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than GB300 NVL72.

1 / 5

Details

AI factory operators commit capital on a megawatt scale, with each megawatt factory costing roughly $60 million. Returns depend on three interdependent factors: earning capacity, useful life, and demand. NVIDIA argues its platform maximizes all three through engineering codesign, long-term hardware viability, and broad workload compatibility.

Productivity: Throughput and Cost

Power is the binding constraint on AI factories, making tokens per second per megawatt the key metric for earning capacity. SemiAnalysis AgentX data indicates that NVIDIA Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than NVIDIA GB300 NVL72, and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model. These gains result from extreme codesign across models, software, compute, networking, and memory.

Durability: Extended Hardware Life

Older generations remain economically viable as not every workload requires the newest hardware. The NVIDIA A100 GPU, shipped in 2020, remains in commercial service, with CoreWeave extending bookings for these units through 2029. Market data supports this longevity: Silicon Data reports a six-year-old A100 retains a quarter of its original value, while Ornn Data finds the market pays 80% as much for a five-year A100 rental contract as for a one-month contract. CUDA ensures software compatibility across generations, preventing stranded assets.

Fungibility: Broad Workload Support

NVIDIA GPUs serve as general-purpose accelerators rather than custom ASICs, running diverse AI and non-AI workloads. The platform supports language, vision, biology, physics, and robotics across all development phases. CUDA-X libraries, numbering over 1,000, enable applications from computational lithography to climate modeling. Real-world deployments include Pinterest using 14,000 GPUs for vision language models and Texas A&M University achieving 95-98% utilization on molecular simulations. This versatility ensures factories remain revenue-generating even as specific AI trends shift.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.