根据 2025 年财务申报，Nvidia 计划在未来 5 年投入 260 亿美元（约每年 52 亿美元）用于建置开放权重（open-weight）AI 模型；主管对 WIRED 证实此一先前未报导的投资。此策略意在把 Nvidia 从以晶片与软体堆叠见长的供应商，推向可与 OpenAI、Anthropic、DeepSeek 竞争的「frontier lab」，并透过将模型调校至其硬体来巩固训练晶片的主导地位。
Nvidia 同步释出 Nemotron 3 Super，宣称为目前最强的开放权重模型；其规模为 1,280 亿参数，约与 OpenAI 的最大版 GPT-OSS 同量级。Nvidia 表示该模型在涵盖 10 项基准的 Artificial Intelligence Index 上得分 37，较 GPT-OSS 的 33 高出 4 分；但亦承认多个中国模型分数更高。公司并称其在 PinchBench（测试操控 OpenClaw）中以秘密测试方式排名第 1。
在开放模型版图上，Meta 于 2023 年率先推出 Llama，但 Mark Zuckerberg 近期暗示未来未必全面开源；相对地，DeepSeek、Alibaba、Moonshot AI、Z.ai、MiniMax 等多个中国模型长期免费公开权重，促使全球新创与研究者广泛采用。Nvidia 研究副总裁 Bryan Catanzaro（2011 年加入）指出，公司已完成 5,500 亿参数模型的预训练，并自 2023 年 11 月首发 Nemotron 后，扩展至机器人、气候建模与蛋白质折叠等专用模型；另有传闻 DeepSeek 新模型可能仅以受美国制裁影响的 Huawei 晶片训练，恐推动更多使用者转向 Huawei 硬体，从而使 Nvidia 的开放投资成为美国对应中国开放模型的关键替代方案。
A 2025 financial filing indicates Nvidia plans to spend USD 26 billion over the next five years (about USD 5.2 billion per year) to build open-weight AI models, a move executives confirmed to WIRED. The strategy aims to push Nvidia from a chipmaker with a strong software stack into a frontier-lab-level model developer that can contend with OpenAI, Anthropic, and DeepSeek, while keeping its hardware central by tuning models to Nvidia systems.
Nvidia also released Nemotron 3 Super, which it calls its most capable open-weight model so far, with 128 billion parameters—roughly comparable in scale to the largest version of OpenAI’s GPT-OSS. Nvidia reports a score of 37 on the Artificial Intelligence Index, which aggregates 10 benchmarks, versus GPT-OSS at 33 (a 4-point gap), while noting several Chinese models score higher. It further claims a No. 1 ranking on PinchBench, a new test focused on controlling OpenClaw, based on undisclosed “secret” evaluation.
Open-weight competition is shaped by release policies: Meta opened Llama in 2023, but Mark Zuckerberg has suggested future superintelligence models may not be fully open, and OpenAI’s GPT-OSS is positioned below its best proprietary systems. In contrast, many Chinese models from DeepSeek, Alibaba, Moonshot AI, Z.ai, and MiniMax publish weights freely, driving widespread downstream adoption; Nvidia frames its openness as ecosystem support and a strategic hedge. Bryan Catanzaro, who joined Nvidia in 2011, says the company has completed pretraining a 550-billion-parameter model, and since the first Nemotron release in November 2023 has expanded into specialized models; rumors that an upcoming DeepSeek model was trained exclusively on Huawei chips raise the stakes by potentially shifting experimentation toward Huawei hardware.