NVIDIA enters the AI large model race

If someone asks who the biggest winner is in the AI era, the answer is almost beyond doubt—NVIDIA. With a steady stream of H100 chips in high demand, it’s like a vendor selling shovels during a gold rush: while AI companies around the world fight tooth and nail, NVIDIA quietly rolls in massive profits, with its market value surging skyward. The latest financial filings show that NVIDIA will invest a total of $26 billion over the next 5 years to fully advance the R&D of open-source AI foundation models. This means NVIDIA is no longer satisfied with just selling shovels—it’s about to jump in personally to dig for gold.

Big-Spending Investment

On March 12, according to financial filings submitted by NVIDIA to the U.S. Securities and Exchange Commission (SEC), NVIDIA will invest a total of $26 billion (about RMB 178.8 billion) over the next 5 years to fully advance the R&D of open-source AI foundation models. NVIDIA has also officially begun its strategic transition from a “chip manufacturer” to a “full-stack top AI laboratory.”

According to the plan, NVIDIA’s $26 billion investment is not focused solely on developing a single model. Instead, it will cover the entire industrial chain of open-source AI foundation models. The funding will be rolled out gradually within the next 18 to 24 months, and the first self-developed open-source AI models are expected to be officially released as early as late 2026 to early 2027.

For comparison, this scale of investment far exceeds the $3 billion spent when OpenAI trained GPT-4. And on the technical roadmap, NVIDIA has chosen a “open-weight” “middle path.” This model sits between OpenAI’s fully closed approach and Meta’s fully open Llama series.

Specifically, NVIDIA will make the key parameters (weights) of the open model public, allowing enterprises and developers to download and use them for free, run and fine-tune them on their own devices or private clouds—meeting companies’ needs for data privacy, customization, and cost control. However, the model’s training data and code may not be fully disclosed.

Andy Konwinski, founder of the non-profit Laude Institute focused on promoting AI openness and a computer scientist, characterized NVIDIA’s investment as a milestone signal. “They are at the point of intersection between many open and closed AI projects,” Konwinski said. “This is an unprecedented statement of their belief in openness.”

In addition, industry analysis notes that an open-source strategy also has longer-term commercial significance for NVIDIA. When NVIDIA releases its models, it will publish weights and technical details, making it easier for startups and researchers to modify and innovate on top of its technical foundation. This helps form a developer network around NVIDIA’s hardware ecosystem, further strengthening the stickiness of NVIDIA’s chips in the market.

Comparable to OpenAI

Since NVIDIA released its first Nemotron model in November 2023, it has gradually rolled out specialized models for vertical fields such as robotics, climate modeling, and protein folding. NVIDIA’s Vice President of Research in Deep Learning, Bryan Catanzaro, also revealed that NVIDIA has recently completed the pre-training of a 550-billion-parameter model. In core model R&D, NVIDIA will focus on developing frontier large models that are multimodal and cross-domain, covering multiple directions such as language, code, scientific computing, and agents.

Recently, NVIDIA also introduced a new generation of open large language model, Nemotron 3 Super, designed specifically for enterprise-grade multi-agent systems. The model has a total parameter count of 128 billion (with inference activating only 12 billion). It natively supports a super-long context window of 1 million tokens. Unlike the mainstream API access model, NVIDIA has opened the model weights, the pre-training/post-training datasets, and the full training scheme.

With 128 billion parameters, the scale is broadly comparable to the largest version of OpenAI GPT-OSS. NVIDIA claims that in the AI Index composite score, Nemotron 3 Super scored 37 points, while GPT-OSS scored only 33 points.

Worth noting is that NVIDIA also acknowledges that the scores of some Chinese models are higher than this level. In addition, NVIDIA says that Nemotron 3 Super participated in a new benchmarking test called PinchBench, which specifically evaluates a model’s control ability over OpenClaw. Nemotron 3 Super ranked first in this test.

On the technical side, NVIDIA has disclosed multiple innovative methods used to train this model, covering architectural and training techniques that improve the model’s inference capability, long-context handling ability, and reinforcement-learning response capacity.

Catanzaro said, “NVIDIA is giving far more attention than ever before to open-model development, and we are making a lot of progress.”

On the ecosystem front, NVIDIA has already reached partnerships with major cloud service providers and hardware vendors such as Google Cloud Vertex AI, Oracle Cloud infrastructure, Dell Technologies, HPE, and others. Integration with Amazon AWS Bedrock and Microsoft Azure is also in the works. Software agent companies such as CodeRabbit, Factory, and Greptile, as well as life science institutions Edison Scientific and Lila Sciences, have also announced that they will integrate this model into their agent workflows.

Redefining the Roadmap

For a long time, NVIDIA’s core advantage has been concentrated in chip hardware. Its global market share in AI chips exceeds 80%, but its influence in the AI model layer has been comparatively weaker. Previously, most technical standards and training paradigms for large models were defined by vendors such as OpenAI and Meta.

NVIDIA’s move to develop a top open-source model in-house is essentially intended to define the technical roadmap for AI models from the ground up—making its own hardware architecture and software stack become the de facto standards across the entire AI industry. By driving compute demand through open-source models, the aim is that if Nemotron becomes the mainstream foundation model for enterprise agent AI, then the GPU infrastructure required to run it at scale will still rely heavily on NVIDIA. While pushing openness at the model layer, NVIDIA also consolidates demand lock-in at the hardware layer.

Financial analysts predict that if NVIDIA can successfully capture 10% share in the foundation model market while consolidating its position as a hardware hegemon, this could add as much as $50 billion in extra revenue per year for the company within three years. Bryan Catanzaro said that promoting the development of an open-source ecosystem fully aligns with NVIDIA’s core interests. This large-scale investment is not blind bandwagoning—it is a strategic choice made after long-term industry research and judgment.

Local time on Tuesday, NVIDIA CEO Jensen Huang also published a rare long-form blog post about artificial intelligence. It is the seventh public long article he has published since 2016. The piece systematically explains the underlying logic of the AI industry, and in the article Huang defines AI’s “five-layer architecture.” He said that the AI industry is still in an extremely early stage of development; although the industry has invested thousands of billions of dollars, the true potential of AI has not yet been fully unlocked. In the future, it will still require ongoing investments of tens of thousands of billions to improve the underlying infrastructure.

Huang pointed out that AI has become one of the strongest forces shaping the world today. It is not a single intelligent application or model, but—like electricity and the internet—it is vital infrastructure. It runs on real-world hardware, energy, and economic foundations, can absorb raw materials and convert them into scalable intelligence. In the future, every company will use AI, and every country will build AI infrastructure.

Regarding employment concerns brought about by AI development, Huang believes that AI will not reduce jobs. Instead, it will create a large number of new employment opportunities—especially in infrastructure and skilled trades. The labor required to support building AI infrastructure is extremely large. AI factories need electricians, plumbers, steelworkers, network technicians, installers, operators, and so on. These are high-skill, high-pay jobs that are currently in short supply. AI is filling the massive global labor gaps in jobs such as truck drivers, nurses, and accountants, rather than manufacturing unemployment.