Nvidia's AI chips already make billions, now Jensen Huang wants $2,000 per GPU in annual software fees - TechSpot

This is Nvidia's Intel Inside moment, transposed onto a SaaS layer. The $2,000-per-GPU annual fee re-prices CUDA from a developer utility into a recurring compute toll — a model echoing Qualcomm's modem licensing, scaled by an order of magnitude. On the technical stack, this adds a second cost layer atop already-expensive inference. Upstream, CoWoS packaging gains pricing power as software lock-in justifies premium silicon, further entrenching TSMC's (Taiwan, China) bargaining position in advanced nodes. Regulatory exposure is concrete. A per-unit SaaS fee invites antitrust scrutiny; Qualcomm's DOJ case over modem royalties is the precedent, and the compute tax framing will not land well in Brussels or Washington. Competitively, AMD and Intel will weaponize zero software tax immediately. More critically, the ROI calculus for AWS Trainium, Google TPU, and Microsoft Maia improves measurably — the avoided fee accelerates hyperscaler de-Nvidia-ification. Within 24 months, software revenue will likely cross 15% of total. But the counterforce is equally fast: Triton, TVM, and open-source compiler stacks will attract institutional funding. The real question is not whether Nvidia collects the fee, but whether the ecosystem fractures structurally around it.