The chronology of the last week is not a coincidence; it is a structural design. On October 15, NVIDIA released Nemotron 3.5 Super VL, providing open weights, open datasets, and training recipes to the market. Six days later, on October 21, the company wrapped a week of high-level European policy engagement with a keynote at GTC Berlin. Between these two events, on October 19, the German federal government and NVIDIA signed a Joint Declaration of Intent to formalize an industrial AI ecosystem. The weights were the gift; the infrastructure was the transaction.
Nemotron 3.5 Super VL is a 120-billion-parameter model that NVIDIA claims delivers the fastest task completion and highest accuracy for specialized always-on agents (vendor-reported). The adopter roster is extensive, featuring Scale AI, EY, Skydance, CrowdStrike, Salesforce, Perplexity, and Dassault Systèmes. This list demonstrates how the model functions as a distribution mechanism. By shipping the model openly, NVIDIA ensures that the software layer is ubiquitous, effectively lowering the barrier for enterprises to begin building. This strategy aligns with the broader industry shift toward a free vendor-neutral serving layer, where the value is captured not by the model license, but by the infrastructure required to run it. Unlike the $20M revenue-gate convergence seen elsewhere, NVIDIA’s OpenMDW-1.1 license carries no such restriction because the rent is collected below the license, at the level of the silicon and the rack.
The keynote in Berlin introduced the concept of sovereign intelligence. According to NVIDIA, this involves companies post-training an open model on their own proprietary data to achieve specialized performance. The asymmetry here is clear: the data and the post-trained model are yours, but the factories, the racks, and the power are the rent. As CEO Jensen Huang noted, the infrastructure must exist long before the models emerge. To facilitate this, NVIDIA is deploying DSX PODs – modular containers that can shrink onsite deployment from six months to 14 days. This is the terminal stage of the weights-to-watts shift, where the hardware landlord gives the software away to ensure the demand for the underlying compute remains constant.
Several corrections are necessary to clarify the record. First, the Nemotron Coalition is a standing initiative, not a Berlin launch; it originated at GTC in March 2026. Perplexity appears on the release page as an integrator, not a named Coalition member, and Mistral does not appear on the page at all. Second, the approximately $2 billion European investment figure that circulated prior to the keynote was media expectation, not an NVIDIA number. The company’s own figures confirm nearly $10 billion committed across Europe. Third, claims that Super VL challenges Opus 5 are secondary framing; NVIDIA’s own comparison is to its largest reasoning model at roughly one-fifth the parameter count (vendor-reported). Finally, the shorthand 120B MoE is not on the primary; the release page specifies an open 120-billion-parameter model with no architecture class named.
For builders and deployers, the binding constraint is moving from the license to the watts. The model is the entry point, but the industrial reality is defined by power, racks, and factories. As we explore in our Model Moats analysis, the ability to deploy at scale is the new competitive frontier. While we put four questions to NVIDIA’s ecosystem contact regarding Coalition membership terms, the European package breakdown, the staging of the October 15 release against the October 21 keynote, and non-public context, those questions remained unanswered at publication time. Consequently, we have relied strictly on the provided primary documentation to map this shift.