Nvidia announced that it will invest $26 billion in the next five years to develop open source AI models - the world's largest AI chip company is changing from a "shovel seller" to a "personal gold mine", directly challenging top AI companies such as OpenAI, Anthropic, and DeepSeek.

This plan has been written into Nvidia's 2025 financial documents and confirmed by executives. Nvidia has released Nemotron 3 Super, the most powerful open-source model currently available, and said it outperformed OpenAI's open-source model GPT-OSS in multiple tests.

This move has far-reaching effects.

This marks a fundamental shift in Nvidia's strategy: no longer just providing chips and software tools, but becoming a full-stack AI company integrating chips, software, and large models, competing head-on with the world's top AI laboratories.

| Nemotron 3 Super: Performance approaches the top models

Nvidia's latest open-source large model, Nemotron 3 Super, has 128 billion parameters, which is similar in size to the largest version of OpenAI's open-source model GPT-OSS. In the authoritative "AI Index" comprehensive score, it scored 37 points, higher than GPT-OSS's 33 points - but Nvidia also admitted that some Chinese large models scored higher than it.

This model also took first place in a new test called PinchBench. This test specifically evaluates the ability of an AI-controlled manipulator (called OpenClaw) to demonstrate its outstanding performance in "letting AI do the work."

Technically, Nvidia has disclosed a number of training innovations, such as how to make the model reasoning stronger, handle longer texts, and give smarter answers through reinforcement learning.

Bryan Catanzaro, Nvidia's vice president of deep learning research, said: "We are investing more in open source models than ever before, and progress is very clear. ”

He also revealed that the company has recently completed the pre-training of a 550 billion parameter large model. Since the launch of the first Nemotron model in November 2023, NVIDIA has successively released a number of specialized models for specific fields, including robotics, climate simulation, and protein structure prediction.

Hardware + model, Nvidia's "two-wheel drive" strategy

Nvidia's efforts to develop large models are not only to compete with others whose AI is stronger, but also to promote the development of its own hardware.

Kari Briski, head of the company's generative AI software business, explained: Future AI models will not only be used to run algorithms, but also to "stress test" the entire data center system - including chips, storage and networks. By training and running these large models, NVIDIA can better understand how to design the next generation of hardware.

In other words: the model is the tool, the hardware is the goal.

In addition, Nvidia chooses open source model weights and technical details, and also has long-term plans. This allows startups, universities and developers to easily do secondary development and innovation on its basis, and over time, an active ecosystem around NVIDIA chips will be formed. The more people use it, the more people can't do without its hardware.

As Bryan Catanzaro, VP of Deep Learning Research, said, "It's in our best interest to help the ecosystem grow." He joined Nvidia in 2011 and was one of the key figures in driving the company's transformation from a gaming graphics card manufacturer to an AI chip giant.

Today, Nvidia is using the combination of "open source model + special chip" to firmly consolidate its moat in the AI era.

Industry experts have praised Nvidia's new strategy

The research community generally speaks highly of Nvidia's vigorous investment in open source large models.

Nathan Lambert, a researcher at the Allen Institute for Artificial Intelligence (AI2) and head of the "True Open Source Model" (ATOM) project in the United States, said he is a "die-hard fan" of the Nemotron model and called on the U.S. government to also fund the development of open source AI.

Andy Konwinski, a computer scientist and founder of the nonprofit Laude Institute, called the investment a "milestone event." He said: "Nvidia is standing at the intersection of open source and closed source AI. This large-scale investment is their most powerful statement on the concept of open source. ”

Investment advice: Nvidia's "software premium" is being revalued

For U.S. Stock Investors (NVDA):

Short-term focus: whether the Nemotron ecosystem can attract the adoption of leading startups and cloud vendors;

Long-term logic: $26 billion of investment will translate into stronger customer stickiness and pricing power, supporting the valuation transition from "hardware cyclical stocks" to "platform technology stocks".

The risk is that if the open source community turns to AMD or domestic chip optimization solutions, the ecological effect may be diluted.

Mapping of A-shares/Hong Kong stocks:

Domestic AI chips (Cambrian, Haiguang, Huawei Ascend) are facing greater ecological pressure - computing power alone is not enough, but also has to have a "useful model + tool chain"; Focus on multi-chip compatible model middleware or cross-platform inference framework startup opportunities.

For developers and entrepreneurs:

Nemotron provides a high-performance, commercial, and fee-free base model option that is especially suitable for agent applications, vertical field fine-tuning, and edge AI deployments.

However, we need to be vigilant: after deeply binding the NVIDIA ecosystem, future computing power costs and technical paths may be limited.