NVIDIA to Invest $26 Bn in Open-Weight AI Models Over 5 Years

NVIDIA  plans to spend about $26 billion over the next five years to develop open-weight AI models, Wired reported, citing a 2025 filing. 

According to the report, the company said it will release models with publicly available weights, allowing developers and researchers to access, modify and deploy them. 

NVIDIA said the approach could increase the use of its computing infrastructure while supporting a wider developer ecosystem.

The company believes the wider release of model weights could accelerate work in areas such as robotics, scientific research and industrial applications while keeping developers within its computing ecosystem.

The company’s investment also reflects increasing competition as open models from companies such as DeepSeek and Alibaba Group have gained traction globally by allowing developers to inspect and adapt models.

The plan comes as NVIDIA launched a 120-billion-parameter model on March 11 that runs agent-based AI systems at scale.

“NVIDIA Nemotron 3 Super is designed to handle complex subtasks inside multi-agent systems,” the company said in its blog post. 

The model is available through NVIDIA’s developer platform and on services including Hugging Face and OpenRouter. Cloud providers, including Google Cloud and Oracle Cloud, offer the model through their AI platforms, with support expected on Amazon Web Services and Microsoft Azure.

The model is packaged as an NVIDIA NIM microservice, allowing enterprises to deploy it on-premises or in the cloud.

Nemotron 3 Super uses a hybrid mixture-of-experts architecture that combines Mamba layers with transformer layers. Only 12 billion of the model’s 120 billion parameters are active during inference.

It is built to support autonomous agents performing multi-step tasks such as software development, research and workflow automation. 

It includes a 1-million-token context window, allowing agents to retain long workflow histories and operate across multi-step tasks.

The model also introduces latent mixture-of-experts, which activates four specialist experts at the cost of one during token generation. It uses multi-token prediction to generate multiple words simultaneously during inference.

On NVIDIA’s Blackwell platform, the model runs using NVFP4 precision, which the company said reduces memory requirements and improves inference speed compared with FP8 on the Hopper architecture.

The company said the model was trained using synthetic data generated by reasoning models.

Several companies have begun integrating the model into their systems. Enterprise software companies, including Amdocs, Palantir Technologies, Cadence Design Systems, Dassault Systèmes and Siemens, are deploying and customising the model for workflows in telecom, cybersecurity and semiconductor design.