Nvidia Will Spend $26 Billion to Build Open-Weight AI Models, Filings Show
Nvidia will spend $26 billion over the next five years to build open source artificial intelligence models, according to a 2025 financial filing. Executives confirmed the news, which has not been previously reported, in interviews with WIRED.
The sizable investment could see Nvidia evolve from a chipmaker with an impressive software stack into a bona fide frontier lab capable of competing with OpenAI and DeepSeek. It’s a strategic move that could further entrench Nvidia’s place as the AI world’s leading chip manufacturer, since the models are tuned to the company’s hardware.
Open source models are ones where the weights or the parameters that determine a model’s behavior are released publicly—sometimes with the details of its architecture and training. This allows anyone to download and run it on their own machine or the cloud. In Nvidia’s case, the company also reveals the technical innovations involved in building and training its models, making it easier for startups and researchers to modify and build upon the company’s innovations.
On Wednesday, Nvidia also released Nemotron 3 Super, its most capable open-weight AI model to date. The new model has 128 billion parameters (a measure of the model’s size and complexity), making it roughly equivalent to the largest version of OpenAI’s GPT-OSS, though the company claims it outperforms GPT-OSS and other models across several benchmarks.
Specifically, Nvidia claims Nemotron 3 Super received a score of 37 on the Artificial Intelligence Index, which scores models across 10 different benchmarks. GPT-OSS scored 33—but several Chinese models scored higher. Nvidia says Nemotron 3 Super was secretly tested on PinchBench, a new benchmark that assesses a model’s ability to control OpenClaw, and ranks number one on that test.
Nvidia also introduced a number of technical tricks that it used to train Nemotron 3. These include architectural and training techniques that improve the model’s reasoning abilities, long-context handling, and responsiveness 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.”
Open Frontier
Meta was the first big AI company to release an open model, Llama, in 2023. CEO Mark Zuckerberg recently rebooted the company’s AI efforts, however, and signaled that it might not make future models fully open. OpenAI offers an open-weight model, called GPT-oss, but it is inferior to the company’s best proprietary offerings, not well-suited to modification.
The best US models, from OpenAI, Anthropic, and Google, can be accessed only through the cloud or via a chat interface. By contrast, the weights for many top Chinese models, from DeepSeek, Alibaba, Moonshot AI, Z.ai and MiniMax are released openly and for free. As a result, many startups and researchers around the world are currently building on top of Chinese models.
“It’s in our interest to help the ecosystem develop,” says Catanzaro, who joined Nvidia in 2011 and helped spearhead the company’s shift from making graphics cards for gaming to making silicon for AI. Nvidia released the first Nemotron model in November 2023. He adds that Nvidia recently finished pretraining a 550-billion-parameter model. (Pretraining involves feeding huge quantities of data into a model spread across vast numbers of specialized chips running in parallel.) Nvidia has since released a range of models specialized for use in areas like robotics, climate modelling, and protein folding.
Kari Briski, VP of generative AI software for enterprise, says Nvidia’s future AI models will help the company improve not just its chips but also the super-computer-scale datacenters it builds. “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.
