Nvidia Challenges OpenAI with $26 Billion for Open-Source AI

Nvidia Bets Big on Open-Source AI

Nvidia recently unveiled its ambitions to become a major player in the open-source artificial intelligence sector. According to a filing with the U.S. Securities and Exchange Commission (SEC), the company plans to invest a staggering $26 billion over five years to develop open-source AI models. This ambitious plan is accompanied by the launch of Nemotron 3 Super, a model distinguished by its power and advanced capabilities.

Chinese giants like DeepSeek and Alibaba currently dominate the open-source sector, while American companies such as Meta and OpenAI have opted to restrict their open offerings. Nvidia sees a strategic opportunity in this situation to attract developers to its hardware ecosystem, positioning itself as a credible Western alternative to Chinese open-weight models.

A Strategic Investment for Nvidia

The SEC filing confirms that Nvidia is ready to invest $26 billion over five years to bolster its presence in the open-source AI model sector. This strategic move aims to counter the growing dominance of Chinese open-source models and to cultivate developer loyalty to its hardware ecosystem.

Alongside this announcement, Nvidia unveiled Nemotron 3 Super, an open-weight model boasting 128 billion parameters. In the Artificial Analysis Index benchmark, this model slightly outperforms OpenAI's GPT-OSS and is comparable to Claude 4.5 Haiku from Anthropic, although it still lags behind some Chinese competitors like Qwen3.5 122B A10B. Nvidia employed several technical innovations during training to enhance reasoning capabilities and manage long contexts. Nemotron 3 Super is a hybrid model combining Transformer architecture with Mamba.

Chinese Dominance in Open-Source AI

Nvidia's investment comes at a time when the balance of power in the realm of open AI models is shifting. Meta was a pioneer with its Llama model, but CEO Mark Zuckerberg recently hinted that future models may not be fully open. Meanwhile, OpenAI offers a GPT-OSS that remains less powerful than its proprietary models, and Anthropic does not provide any open models at all.

In the meantime, Chinese companies like DeepSeek, Alibaba, Moonshot AI, and MiniMax are releasing almost all their model weights for free. Although changes are underway, including departures within Alibaba's Qwen team, Chinese models continue to be the best open alternative for many applications. However, recent benchmarks indicate that the practical gap with the best Western models is often greater than what previous benchmarks suggested. Despite the effectiveness of Chinese models, widespread adoption in Western industry has not materialized.

Nvidia and the Hardware Opportunity

In January 2025, DeepSeek shook the market with a high-performing open-source model that challenged the lead of Western AI labs and the amount of hardware needed to compete. A new model from DeepSeek, trained exclusively on chips from Chinese manufacturer Huawei, could trigger another upheaval. If confirmed, more companies might turn to Huawei hardware, especially in China.

Reports indicate that DeepSeek is also using sanctioned Nvidia Blackwell GPUs for its training. Under pressure from the Chinese government, DeepSeek attempted to train on Huawei chips but encountered technical issues, including unstable performance and an immature software toolchain. Meanwhile, Nvidia has received permission to export more powerful AI chips to China again, although Chinese companies are seeking to avoid renewed dependency.

By releasing its own open models optimized for Nvidia hardware, the company creates a counterbalance and offers an alternative for Western companies. Those who adopt Nemotron and its associated models remain within the Nvidia ecosystem. The company is also targeting markets where major AI labs are absent, such as robotics and other edge AI applications.

Bryan Catanzaro, Vice President of Applied Deep Learning Research at Nvidia, told WIRED: "We are an American company, but we work with companies around the world. It is in our interest to make the ecosystem diverse and strong everywhere."

Nvidia has already pre-trained a 550 billion parameter model and has released specialized models for fields such as robotics, climate modeling, and protein folding.

Kari Briski, Vice President of Generative AI Software, highlighted another strategic dimension: the models serve to test the resilience of data centers at the scale of Nvidia's supercomputers and to advance its hardware roadmap. "We build them to stretch our systems and test not only computation but also storage and networking, and to develop our hardware architecture roadmap," she told WIRED.