Comment on FR Doc # 2026-12205

NVIDIA appreciates the opportunity to comment on Notice-MVAC-2026-01 governing the safeguarding of Government Data within large language model systems, and to participate in GSA’s July 14, 2026 public listening session. We write to ensure that the framework protects, rather than forecloses, open-weight AI models — and we have attached specific, insertable regulatory language to that end. Since July 24, 2026, ten days after the public listening session for 552.239-7001, NVIDIA has joined with more than 230 other American technology companies and organizations — including Microsoft, Meta, Dell, IBM, Palantir, Hugging Face, the Linux Foundation, and Mozilla — in an open letter titled “Open Weights and American AI Leadership.” That letter, shared publicly by NVIDIA’s CEO Jensen Huang, makes a case directly relevant to this rulemaking: open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable national sovereignty. It draws an explicit parallel to the open-source software movement of the 1980s — the same lineage that produced the Linux systems the federal government now runs across its most sensitive missions. The concern underlying that letter squarely applies here. As drafted, the LLM Developer flowdown attaches contractual obligations to any party that publishes model weights, including open-weight model publishers that never receive or process Government Data and have no relationship with the ordering agency. Because such a publisher cannot accept, perform, or attest to those obligations, the clause would in practice exclude open-weight models from federal procurement altogether — the opposite of the safeguarding outcome GSA intends, and contrary to the national interest that over 150 leading AI companies articulated (see attached letter). The attached recommendations resolve this through a single principle: obligations should follow the data, not the model’s authorship. They preserve every safeguarding requirement in the clause and reassign each to the party that actually has access to and processes Government Data — the system operator and integrator — while recognizing that a self-hosted open-weight model enforces the clause’s core protections through architecture rather than promise. Open-weight models — those that openly publish their trained weights, giving the Government full access to inspect, customize, fine-tune, and self-host — are the most transparent, auditable, and sovereign AI systems available to federal agencies. Models that also publish their training data represent the fullest expression of this openness. NVIDIA urges GSA to ensure this clause protects open-weight models, and we welcome the opportunity to provide technical detail on any of our recommendations.