Nvidia released Nemotron 3.5 Transcribe, an open-weight model for multilingual audio transcription and translation, alongside Granary v2, an open-source speech dataset designed to train and refine speech recognition systems.
The strategy is deliberate: by providing free, open-weight models, Nvidia reduces friction for developers and enterprises to build AI applications—which then run on Nvidia's high-margin GPUs. Open-weight architecture, which allows inspection and modification of the underlying AI model, contrasts with proprietary closed-source alternatives and positions the company as a foundational infrastructure provider rather than a software vendor.
The transcription and translation capabilities address a concrete market need. Enterprises operating across linguistic markets can deploy Nemotron 3.5 Transcribe for customer service, content localization and data analysis without licensing costs or vendor lock-in at the model layer. That apparent openness masks Nvidia's real advantage: any company that builds on these models still requires Nvidia hardware to train, fine-tune and serve them at scale.
Granary v2's open-source collaborative model accelerates adoption. Researchers and developers contribute improvements, widening the dataset and making the model more useful—which increases its installed base and, critically, the number of production workloads running on Nvidia infrastructure.
Nvidia has built its competitive position by controlling both the hardware and the software layer. CUDA, the company's parallel computing platform, created a ten-year moat by making it prohibitively expensive for enterprises to port workloads to rival chips. Nemotron and Granary extend that logic into AI applications: the easier Nvidia makes it to start building, the harder it becomes to leave.
