Nvidia redefines AI factory economics with productivity, durability, flexibility

Nvidia is reframing the economics of AI factories around three principles — productivity, durability and flexibility — designed to maximize return on facilities that cost roughly $60 million per megawatt. The company argues that the combination of maximum token throughput per watt, an economic lifespan stretching well beyond original projections, and the ability to run any workload type determines whether a factory ever recovers its capital.

Productivity is measured in tokens per second per megawatt, because power is the hard bottleneck and that number sets the revenue ceiling. According to SemiAnalysis AgentX data, Vera Rubin NVL72 systems deliver 30 times more throughput per megawatt than GB300 NVL72, and up to 45 times lower cost per million tokens on DeepSeek V4 Pro. The leap comes from extreme co-design across the entire stack — models and workloads, software, compute, networking and memory developed together. The direct result: more revenue inside a fixed power envelope, and wider margins on every token.

The obvious question is whether such a dramatic price drop shrinks the need for compute. Nvidia's answer is no. Cheaper tokens make more use cases economically viable, and those use cases consume more tokens than the efficiency gains save. It is a classic Jevons paradox: efficiency expands the market instead of contracting it. Flexibility (Fungible) completes the picture. The same systems run every AI workload — training, inference, fine-tuning — plus non-AI workloads, through CUDA-X libraries and a standard architecture available in a validated reference design.

Durability is proven by the A100, launched in 2020 and still in active commercial service six years later. CoreWeave has extended orders for that generation through 2029. Every major operator has repeatedly pushed out depreciation schedules — the estimate of when hardware stops earning — and that line keeps moving forward. A September 2026 Sprout analysis, "The Productive Life of a Data Center GPU," documents the trend across all key players.

Independent secondary-market data reinforces the claim. Barkr estimates a 5–6 year useful life for an 8-GPU H100 system and 9–10 years for GB300 NVL72, based on resale prices. Silicon Data shows a six-year-old A100 still worth a quarter of its original cost, while a five-year depreciation schedule had written it to zero more than a year ago. Ornn Data finds the market willing to pay 80% of a monthly contract price for a five-year contract on the same GPU — a clear signal of confidence in continued throughput.