Nvidia plans to commit $26 billion over the coming five years to develop open-source artificial intelligence models, as detailed in a 2025 financial report. Company executives validated this information, which had not been previously disclosed, in discussions with WIRED.
This substantial capital outlay could enable Nvidia to evolve from a semiconductor producer with an impressive software ecosystem into a legitimate pioneering laboratory capable of vying with entities like OpenAI and DeepSeek. It represents a calculated maneuver that could further solidify Nvidia’s position as the AI world’s preeminent chip manufacturer, given that these models are specifically optimized for the company’s hardware.
Open-source models are those for which the weights or parameters governing a model’s behavior are publicly disseminated—sometimes alongside details of its architectural design and training methodology. This arrangement permits anyone to download and execute them on their personal machine or within cloud environments. In Nvidia’s particular instance, the corporation also discloses the technical advancements involved in constructing and training its models, thereby facilitating the adaptation and further development of the company’s innovations by startups and researchers.
On Wednesday, Nvidia additionally unveiled Nemotron 3 Super, its most proficient open-weight AI model released to date. This novel model incorporates 128 billion parameters (a metric indicating the model’s scale and intricacy), rendering it roughly comparable to the largest iteration of OpenAI’s GPT-OSS, although the company claims it surpasses GPT-OSS and other models across various evaluation criteria.
Specifically, Nvidia asserts that Nemotron 3 Super achieved a rating of 37 on the Artificial Intelligence Index, a system that assesses models across 10 distinct benchmarks. GPT-OSS scored 33—however, several Chinese models attained higher scores. Nvidia states that Nemotron 3 Super underwent covert evaluation on PinchBench, a new benchmark designed to gauge a model’s capacity to control OpenClaw, and secured the top position on that assessment.
Nvidia also presented a range of ingenious methods it employed to train Nemotron 3. These encompass structural and developmental techniques that enhance the model’s cognitive abilities, extended context processing, and adaptability to reinforcement learning.
“Nvidia is taking open model development much more seriously,” says Bryan Catanzaro, VP of applied deep learning research at Nvidia. “And we are making a lot of progress.”
Unrestricted Horizon
Meta was the initial major AI firm to launch an open model, Llama, in 2023. CEO Mark Zuckerberg, however, recently revitalized the company’s AI initiatives and indicated that it might not make future models entirely accessible. OpenAI offers an open-weight model, designated GPT-oss, but it is less capable than the company’s optimal proprietary solutions and not well-suited for modification.
The foremost US models, originating from OpenAI, Anthropic, and Google, are accessible exclusively through cloud platforms or via a conversational interface. In stark contrast, the parameters for numerous leading Chinese models, from DeepSeek, Alibaba, Moonshot AI, Z.ai, and MiniMax, are publicly disclosed and offered without charge. Consequently, many emerging companies and researchers globally are presently developing solutions based on Chinese models.
“It’s in our interest to help the ecosystem develop,” says Catanzaro, who joined Nvidia in 2011 and helped pioneer the company’s transition from producing graphics cards for gaming to creating silicon for AI. Nvidia launched the initial Nemotron model in November 2023. He further mentions that Nvidia recently concluded the pretraining phase for a 550-billion-parameter model. (Pretraining entails ingesting enormous volumes of data into a model distributed among a multitude of dedicated processors operating concurrently.) Nvidia has subsequently introduced an array of models tailored for applications in domains such as robotics, climate simulations, and protein folding.
Kari Briski, VP of generative AI software for enterprise, asserts that Nvidia’s forthcoming AI models will assist the company in enhancing not only its chips but also the vast, supercomputer-scale data processing facilities it constructs. “We build it to stretch our systems and test not just the compute but also the storage and networking, and to kind of build out our hardware architecture roadmap,” she says.
{content}
Source: {feed_title}