Nvidia spent the AI boom selling the machinery everyone needed to build models.
Now, it increasingly wants to build the models, too.
The company is developing Nemotron 4, a new family designed to compete with leading open-weight AI systems. According to reporting from The Information cited by Reuters, the largest planned version will contain at least one trillion parameters. Final training is not yet complete, and release could come as early as late autumn. �
Reuters
Nvidia has confirmed work on Nemotron 4 generally but has not independently confirmed every reported specification. �
Reuters
That qualification matters.
Still, Nvidia’s direction is fascinating.
Nvidia is climbing the AI Stack
The company already controls a hugely important layer:
compute.
It is now expanding into:
models,
model routing,
security,
development software,
infrastructure finance.
Nvidia recently released Nemotron 3.5 Lightning for tasks such as code review and security monitoring and introduced NeMo Switchyard, an open-source system for routing jobs between models. �
Reuters
This is no longer simply a chip company.
It is increasingly constructing the operating infrastructure surrounding artificial intelligence.
Does One Trillion Parameters Mean One Trillion Intelligence Points?
No.
Parameter count is not an IQ score.
Architecture, training, data quality, reasoning methods, and inference efficiency can matter enormously.
A smaller system may outperform a larger one on particular tasks.
The interesting question is, therefore, not merely Nemotron’s size.
It is whether Nvidia can combine a competitive open model with its enormous hardware and developer ecosystem.
Mary Chuks’ Perspective
Nvidia understands the scaler strategy beautifully.
First, sell everyone the kitchen.
Then, provide the ingredients.
Then, provide the recipe.
Eventually, you realise Nvidia is standing beside you, asking whether you’d like them to cook dinner too. 🤣
The strategic danger for competitors is vertical integration.
Conclusion
Nemotron 4 signals that the boundaries of the AI industry are disappearing.
Model companies build chips.
Chip companies build models.
Cloud companies finance data centres.
Everybody is quietly becoming everybody else.
Primary source: Reuters, 11 August 2026. �
Reuters