$16K per Node: What Per-GPU Software Licensing Adds to a Self-Hosted AI Bill
A report from November 2026 puts Nvidia's enterprise software licensing at roughly $2,000 per GPU. On an eight-GPU node that becomes $16,000 per node per year — about 6% of a $280,000 box, but closer to 12% of the node's annual run cost once hardware amortization is included. At full availability it is $0.23 per GPU-hour. This short walks through the arithmetic: the per-node multiplication, the shift in break-even duty cycle from roughly 65% to roughly 74% against cloud on-demand, why idle and warm-standby capacity still bill, and why committed-use cloud rates push the crossover past 100% duty cycle, where the node never pays back with or without the fee. The post is explicit about what is confirmed and what is assumed: the $2,000-per-GPU figure comes from the report, while the annual, per-physical-GPU, recurring framing is the author's assumption. The report does not define the licensing unit, the term, or whether the fee is additive to existing per-GPU subscriptions such as NVIDIA AI Enterprise — each of which changes the answer by more than an order of magnitude. Full cost model, editable worksheet and measurement plan: https://hackyjs.com/posts/16k-per-node-what-per-gpu-software-licensing-adds-to-a-self-hosted-ai-bill