Nvidia presents chips for « AI factories »

The semiconductor giant Nvidia wants to meet the rapidly growing need for computing power for artificial intelligence with a new generation of its chips. The new system called « Vera Rubin » is scheduled to come onto the market in autumn 2026, said Nvidia boss Jensen Huang at the GTC developer conference.

Rubin and the further development of the current Blackwell platform for this year should drastically reduce the costs for the operation of AI software compared to previous technology.

Nvidia chips became a key technology for artificial intelligence. First of all, the systems around the world were used primarily for training applications with artificial intelligence. Tech giants such as Google or the Facebook group Meta fill entire data centers-but also AISTPUPS such as the Chatgpt-FinderFIRMA OPENAI. This position had the NVIDIA business grow explosively in recent years.

The next generation of chip is named after Vera Rubin – an American astronomer who made important discoveries in dark matter.

« AI factories » and robot

The Nvidia boss speaks of data centers as « AI factories » in which the computer performance is for AI. « Every industry that makes something will have two factories in the future, » said Huang. One will manufacture the physical products as before – and the second will deliver the software for it.

Huang presented a new platform at the in-house GTC conference to accelerate the development of humanoid robots. Nvidia works with Disney and Google's Ki company Deepmind in the project called Isaac Groot N1. In the development of technology for autonomous driving, Nvidia won the US car giant General Motors as customers for his computer systems and software.

100 times more computing power than expected

At the same time, Huang tried to dispel concerns from investors that the world could get by with less AI in the future – and that the expectations of the future business of Nvidia could also be too high.

Overall, the world is about to generate answers with the help of artificial intelligence instead of calling up stored answers. In particular, the new AI models, which can build a chain of argument step by step, are hungry for performance. As an example, Huang demonstrated how that Chinese model Deepseek R1 150 times more computing power than traditional AI software needed to calculate the correct seating arrangement at a wedding. The conventional model also failed due to the task.

In total, you need 100 times more computing capacity than accepted a year ago, said Huang. It shouldn't have been a coincidence that he selected Deepseek for demonstration. R1 is said to have been trained with significantly less computing effort than previous AI models. That had one a few weeks ago Course fall triggered of the Nvidia share. However, Huang argues that the actual need for computing power will not arise during training, but when the answers were created.

In the meantime, investors were not quite convinced: the Nvidia share in the US trade with a minus of 3.43 percent.