NVIDIA is fully developing the Nemotron 4 open-source large model to drive chip demand while directly facing customer competition

NVIDIA is intensifying its efforts in the open-source arena, investing resources to develop a significant large model in-house. The company hopes to drive hardware demand through this model, but this move also means it may compete with its own customers and partners.

Several individuals involved in the Nemotron project revealed that NVIDIA plans to create the largest foundational model in the Nemotron 4 series, aiming for performance that rivals the world's top open-source large models. The previous generation flagship large model had 570 authors, and employees indicated that the number of personnel involved in the development of Nemotron 4 will further expand. A former employee stated, "At this stage, everyone hopes to be involved."

Recently, NVIDIA has continuously launched multiple open-source models, and this development builds on that foundation. On Tuesday, NVIDIA released the Nemotron 3.5 Lightning lightweight model, focusing on efficient and high-speed operation of intelligent agents. The company also introduced free model routing software to help enterprises quickly build model scheduling tools—these tools can allocate different AI tasks to the most suitable and cost-effective models.

Multiple employees from the Nemotron project disclosed that the flagship Nemotron 4 will have at least trillions of parameters. The parameter count refers to the adjustable units during the model's learning process, and this scale is approximately twice that of NVIDIA's current flagship model, Nemotron 3 Ultra (released in June). Even with trillions of parameters, the model's size is still smaller than that of leading open-source large models in the U.S.; however, NVIDIA places great importance on model compression technology, which may allow smaller models to achieve better performance.

NVIDIA is increasing its investment in training computing power for self-developed models, highlighting its determination to advance its open-source strategy. NVIDIA is leasing AI servers from cloud vendors that purchase its chips to obtain computing resources. As of April, the total amount of long-term cloud service procurement agreements signed by the company has increased to $28 billion (contracts lasting until early 2031), approximately three times the scale disclosed a year ago. Some computing resources are also allocated to other R&D projects.

The strong push for the Nemotron project places NVIDIA in a delicate position: effectively competing with a number of open-source startups that have received its investment, as well as leading AI laboratories that are its most important customers. OpenAI has consistently been one of the largest sources of demand for NVIDIA chips, and NVIDIA has invested $30 billion in this AI laboratory.

Although the performance of the flagship Nemotron 4 is expected to fall short of the closed-source cutting-edge models from OpenAI and Anthropic, open-source models are increasingly being adopted by more enterprises due to their cost advantages, taking on certain business scenarios.

NVIDIA has also made significant investments in several domestic open-source companies in the U.S., which are dedicated to promoting self-developed large models, including Reflection AI and Thinking Machines Lab.

More than a dozen current and former employees, NVIDIA partners, and Nemotron users indicated that NVIDIA's increased focus on open-source is based on the core logic that a rich open-source ecosystem can continuously drive GPU demand. Currently, the demand for NVIDIA chips is highly concentrated among a few leading laboratories and cloud vendors, such as OpenAI, Microsoft, and SpaceX, some of which have already begun developing their own AI chips NVIDIA's approach is that if a large number of startups and traditional enterprises can rely on NVIDIA to optimize and adapt hardware with high-cost performance open-source models for AI development, or accelerate innovation across the entire open-source track, NVIDIA will benefit from it.

Anastasios Angelopoulos, CEO of the AI model evaluation organization Arena, stated: "Whichever company produces an excellent open-source model, NVIDIA is the winner."

Boosting Competition in the Open-Source Track

An NVIDIA executive stated that the company believes that self-developed large models can stimulate industry competition, drive more open-source model development, and ultimately increase GPU demand. NVIDIA has established the "Nemotron Alliance," inviting several American open-source development companies to join, where all parties jointly provide training data, research ideas, and collaborate on developing the next generation of models.

Kari Briski, NVIDIA's Vice President of Generative AI, stated in an email: "NVIDIA's continued investment in the Nemotron project is because we believe that companies of all types and from all countries need easily accessible cutting-edge open-source models to strengthen security, accelerate innovation, and create a technology foundation that can be continuously iterated and relied upon for the long term."

The specific launch date for Nemotron 4 has yet to be determined. Several employees stated that NVIDIA has finalized the pre-training data and infrastructure direction for the flagship Nemotron 4, but precise parameter specifications and the release date have not been established, and the final complete training process has not yet started, with the training cycle potentially lasting several months. Two employees expect the model to be completed by late autumn this year; others believe it may be later.

Brian Catanzaro, NVIDIA's Vice President of Deep Learning Research, mentioned in a podcast in January that the layout of the Nemotron project "is crucial for the company's long-term future."

NVIDIA's large models have already been implemented by some companies, with a typical case being Palantir. This company announced in June a partnership with NVIDIA to deploy the Nemotron series models for U.S. government clients.

However, NVIDIA's model overall performance has not yet reached the industry's top tier. As of now, NVIDIA's strongest model, Nemotron 3 Ultra, ranks second in agent capability and comprehensive intelligence metrics among U.S. open-source models in multiple evaluations by Arena and Artificial Analysis, only behind the newly released Inkling from Thinking Machines; there is still a gap compared to the top U.S. open-source models, with a global comprehensive ranking outside the top 40.

Due to constraints on overall corporate funding allocation, NVIDIA's investment in large model development has an upper limit. The budget for NVIDIA's cloud services corresponds to an investment scale of $7 billion for the fiscal year ending in January 2028. Although this is far less than OpenAI and Anthropic, it has already exceeded the total funding amount of most leading open-source laboratories since their establishment. An employee stated that NVIDIA is striving to allocate more computing resources for Nemotron Members of the Nemotron alliance include Reflection, Cursor, Thinking Machines, and Mistral. These companies are advancing their own open-source projects while providing training data, evaluation support, and model design ideas for Nemotron 4. According to informed sources involved in the collaboration, the startup Prime Intellect is providing 300,000 simulated environments for model training. The AI code startup Cognition is also a member of the alliance (the collaboration has not yet been officially announced), and the two parties are negotiating for Cognition to provide code training data to NVIDIA.

Partners have their own considerations for joining the alliance. For companies developing their own large models, NVIDIA's push for open-source can enhance market attention on the open-source track and expand the industry pie; other companies hope to participate in co-building models to directly use the results in the future, saving on high training computing costs.

Vincent Weiser, CEO of Prime Intellect, believes that the significance of forming the alliance is to unite industry forces to prevent the market from being monopolized by a single "ultimate large model," rather than competing to see who can create the strongest open-source model