NVIDIA Vera Rubin Targets 30x AI Factory Throughput per Megawatt
NVIDIA says Vera Rubin NVL72 delivers 30x more throughput per megawatt than GB300, as CUDA extends GPU life, expands demand and cuts token costs.
Summary
NVIDIA argues that AI factory returns depend on earning capacity, useful life and demand, with each megawatt costing roughly $60 million and power limiting output. SemiAnalysis AgentX data shows Vera Rubin NVL72 delivering over 30x the throughput per megawatt of GB300 NVL72 and up to 45x lower cost per million tokens for DeepSeek V4 Pro. NVIDIA attributes those gains to full-stack codesign spanning models, software, compute, networking and memory, while CUDA supports workloads across hardware generations.
A100 GPUs shipped in 2020 remain commercially active, and CoreWeave extended bookings through 2029. Sprout’s September 2026 analysis found every major operator extending server depreciation schedules, while Microsoft ran its V100 fleet for 8.4 years against a six-year book life. Barkr estimates five to six useful years for eight-GPU H100 systems and nine to 10 years for GB300 NVL72. Silicon Data values six-year-old A100s at 25% of original cost, while Ornn Data says five-year A100 rentals command 80% of one-month pricing. More than 1,000 CUDA-X libraries and 10 million developers support AI, simulation, graphics and scientific computing. Deployments include Lilly’s 1,016-GPU cluster, Pinterest’s 14,000 GPUs, Revolut, Runway, Texas A&M’s 95% to 98% utilization across 26 projects and seven institutions, Dassault Systèmes at Wichita State and Lucid Motors, and Unilever cutting digital-twin imagery costs by half. Jensen Huang will discuss AI factories at GTC Berlin on October 21, 2026, at 11 a.m. CEST.
Positives
- Vera Rubin NVL72 delivers over 30x GB300 NVL72 throughput per megawatt and up to 45x lower DeepSeek V4 Pro token costs.
- CoreWeave extended bookings for A100 units introduced in 2020 through 2029, indicating continued demand for older GPUs.
- Barkr estimates GB300 NVL72 systems can remain economically useful for nine to 10 years.
- More than 1,000 CUDA-X libraries and 10 million developers broaden NVIDIA infrastructure beyond AI into simulation, graphics and scientific computing.
- Texas A&M sustains 95% to 98% utilization across 26 projects and seven institutions.
- Unilever’s digital-twin product imagery cuts production costs by half compared with photo shoots.
Risks & concerns
- Each megawatt of AI factory capacity costs roughly $60 million, requiring operators to establish a credible return before committing capital.
- Power is the binding constraint, making throughput per megawatt decisive for revenue capacity.
- High output cannot compensate if operators sell only part of their available token capacity.
- Factories optimized for one workload risk losing utilization and revenue when demand changes.
- Server depreciation schedules remain estimates, and accounting life is only a conservative proxy for physical operating life.