Nvidia's On-Prem Pivot: A Direct Play for the ROI-Conscious Enterprise
AI-generated image · US National Wire
By launching open-weight models for data centers and AI PCs, Nvidia is targeting organizations weary of cloud dependencies and unpredictable API costs.
For the enterprise leader, the allure of frontier AI is often dampened by two persistent frictions: the unpredictable cost of API scaling and the security risks associated with sending proprietary data to the cloud. Nvidia is now positioning itself as the primary alternative to this trade-off.
As TechCrunch first reported, Nvidia has launched a new suite of open-weight models designed specifically for edge computing and always-on agents. The centerpiece for the enterprise is the **Nemotron 3.5 Super VL**, a 120-billion-parameter model. While it utilizes fewer parameters than the Nemotron 3 Ultra, Nvidia asserts that the Super VL maintains similar accuracy in tool-use and instruction-following tasks.
From an operational lens, the value proposition here is ownership. In a blog post, Nvidia’s vice president of Open Models and Software Kari Briski explained that organizations can deploy Super VL on their own infrastructure and post-train it using their own data, unlike the experience with closed frontier models. This allows companies to own the resulting intellectual property rather than renting it via a third-party API.
Nvidia claims that when post-trained for specific agentic tasks—such as medical insurance claims, financial accounting, security attack resilience, scientific experimentation, and document intelligence—the Super VL outperforms competing models including **OpenAI's GPT 6.1 Sol** and **Anthropic's Opus 5, Luna, and Astra**. To validate these claims, Nvidia highlighted a roster of early adopters including **Salesforce, Perplexity, Crowdstrike, Ramp, EY, and Skydance**.
Beyond the data center, Nvidia is pushing into the local hardware market to capture the "privacy-weary" segment of the workforce. TechCrunch reports the launch of two lightweight **Bolt** models designed to run locally on workstations powered by the **DGX Spark** or new AI PCs equipped with the **Nvidia RTX Spark** chip. These machines are hitting the market now through partners including **Dell and Microsoft**.
Of these local models, the **Nemotron 3.5 Bolt 74B-A9B 24GB** (a 74-billion parameter model) is built for multistep reasoning, web browsing, and tool calling. A smaller model, the **Bolt E4B**, is intended for conversational agents and will be released later this month. As Briski noted, this local approach allows users to delegate projects without sending workloads to the cloud.
This strategic shift comes as proprietary AI labs—who are simultaneously Nvidia's customers and investments—diversify their hardware dependencies. TechCrunch notes that these labs are increasingly turning to **Cerebras** and **AWS’s Trainium**, or developing their own in-house silicon, prompting Nvidia to lean harder into the enterprise demand for self-hosted, open-weight alternatives.