Sovereignty Over Subscriptions: The TCO Play in Nvidia's New Open Models
AI-generated image · US National Wire
By prioritizing open-weight models for on-prem and edge deployment, Nvidia is offering enterprises a path away from proprietary API dependencies toward lower total cost of ownership.
For the enterprise leader, the primary friction point in scaling AI isn't just raw capability—it's the compounding cost and risk of proprietary API dependencies. Nvidia's latest release of open-weight models signals a strategic pivot toward reducing this total cost of ownership (TCO) by enabling sovereign, on-prem agentic workflows.
As TechCrunch first reported, Nvidia has introduced the Nemotron 3.5 Super VL, a 120-billion-parameter model specifically engineered for enterprise agents. While the model utilizes fewer parameters than the Nemotron 3 Ultra, Nvidia asserts it maintains similar accuracy in tool-use and instruction-following. The ROI here is not merely technical; it is operational. Kari Briski, Nvidia’s vice president of Open Models and Software, noted that these models allow organizations to post-train on their own data and deploy on their own infrastructure, meaning they "own what they build" rather than renting intelligence from a closed frontier lab.
This shift toward sovereignty is critical as enterprises seek to avoid the "black box" nature of proprietary systems. Nvidia claims that when post-trained for specific agentic tasks—including security attack resilience, medical insurance claims, scientific experimentation, financial accounting, and document intelligence—the Super VL outperforms competing models such as OpenAI's GPT 6.1 Sol and Opus 5, as well as Astra and Luna from Anthropic. TechCrunch reports that a slate of heavy-hitter companies, including Salesforce, Perplexity, Crowdstrike, Ramp, EY, and Skydance, are already utilizing Super VL.
Beyond the data center, Nvidia is pushing the TCO reduction to the edge. The company launched two lightweight Bolt models designed to run locally on workstations powered by DGX Spark and AI PCs equipped with the Nvidia RTX Spark chip. These machines are hitting the market via partners including Dell and Microsoft.
The hardware-software integration serves two distinct operational needs:
1. **High-Reasoning Local Tasks**: The Nemotron 3.5 Bolt 74B-A9B 24GB (a 74-billion parameter model) is designed to balance speed with reasoning for web browsing, tool calling, and multistep reasoning. 2. **Conversational Utility**: The Bolt E4B, arriving later this month, is aimed at conversational agents.
From an ops lens, the value proposition is clear: by moving workloads from the cloud to local hardware, users can delegate projects without transmitting sensitive data to external servers. This addresses the privacy concerns of knowledge workers and programmers while eliminating the recurring per-token costs associated with proprietary APIs.
This aggressive push into open models comes at a precarious time. TechCrunch notes that proprietary AI labs—which are often both Nvidia's customers and investments—are increasingly diversifying their hardware needs, turning to Cerebras and AWS’s Trainium or developing in-house chips. By empowering the enterprise to own its model weights and infrastructure, Nvidia is positioning itself as the foundational layer for the sovereign AI era, shifting the value from the model subscription to the underlying compute.