NVIDIA's AI Factory Claims Called into Question by Independent Sources
NVIDIA's recent blog post on AI-factory returns cites five sources to support its claims. The post, titled 'Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,' argues that each megawatt factory costs roughly $60 million and that the company's platform maximizes return on investment through productivity, durability, and fungibility.
However, a closer look at the sources reveals some discrepancies. SemiAnalysis's data shows that Vera Rubin NVL72 systems deliver up to 30x higher throughput per megawatt than NVIDIA GB300 NVL72, but this advantage depends on both interactivity target and serving engine. The company also notes that the engine label is essential when quoting high-interactivity gain.
Sprout's chart shows that major operators have extended server life, with Microsoft running several GPU generations for 6.9 to 8.8 years. Ornn Data finds that the market pays 80% as much to rent an A100 GPU on a five-year contract as on a one-month contract.
The figures and claims made by NVIDIA in its post are not independently verified, and some sources could not be accessed due to technical issues. Additionally, several statements made by NVIDIA, including the $60 million per megawatt figure and the claim that CoreWeave extended bookings on 2020 units through 2029, have not been checked.
It's worth noting that nothing in this piece is investment advice, and the figures mentioned are attributed to their respective sources. The post does not establish whether an AI factory earns a good return.