Nvidia discusses $2,000-per-GPU annual fee for software that 'runs agents on AI factories'

This isn't a software license—it's a toll on the agent economy. By rebranding CUDA from a programming interface into a runtime OS under the "AI factory" banner, Nvidia is extracting recurring margin from what was a one-time hardware sale. The Oracle-1990s playbook, amplified 100x: 100K GPUs at $2K/year equals $200M in near-zero-marginal-cost revenue. The technical implication is structural. Agent orchestration, memory scheduling, and task routing get absorbed into Nvidia's stack. vLLM and LangChain become thin wrappers. AMD's ROCm and Intel's oneAPI lack agent-native primitives; their likely counter—free inference software—fails because ecosystem gravity isn't a spec-sheet argument. The 12-24 month trajectory: expect tiered per-agent or per-token pricing to replace flat GPU fees. The existential threat isn't antitrust scrutiny. It's AWS Trainium and Azure Maia building in-house agent runtimes that bypass the toll entirely. Nvidia's window to lock the standard is roughly 18 months. Miss it, and "AI factory" collapses into commoditized compute.