NVIDIA CEO: Someday we will have 1B robotic cars on the road

Nvidia Executive Director Jenson Huang predicts that one day we will have a billion cars on the road and all of them will be robotic cars.

It sounds like science fiction, but as Huang said before, “I’m science fiction.” He made the comments in a conference conversation with NVIDIA FYQ4 profit analysts ending January 26, 2025 (here’s our Report of profit). NVIDIA shares are up to date with half a percent up to $ 130.72 per share in trade in hours.

Colette Kress, EVP and CFO said in the conference conversation that business with the data center increased 93% compared to a year and 16% in succession, while the Blackwell ramp began and sales of a bunker chip also increased. Blackwell sales have exceeded NVIDIA’s expectations, she said.

“This is the fastest product ramp in the history of our company, unprecedented in its speed and scale,” Cress said. “Blackwell manufacturing is in full equipment in multiple configurations and simply increasing supplies,
Expanding customer reception. Our Q4 data center calculated the revenue jumped 18% consecutively and more than 2 times a year. Customers are competing to scathing infrastructure for the next generation of vanguard models and unlock the next level of AI capabilities. “

It will be customary with Blackwell to start with 100,000 graphics graphics units (GPU) or more, Cress said. The shipments have already started for multiple infrastructures of this size. Following the training and customization of the model, they nourish the search for NVIDIA infrastructure and software such as developers and corporators leverage techniques such as fine tuning, strengthening of training and distillation of model adaptation. Hugs alone publishes over 90,000 derivatives created by the Llama Foundation model.

The scale of the following training and personalization of the model is massive and can collectively require more orders than pre -training, Cress said. And the search for conclusions is accelerated, driven by the scales of time and new models for reasoning such as Openai O3, Deepseek and others. Cress said China expects to increase consistently, and Huang said China is expected to be the same percentage as in Q4. This is about half of what was before the control of the export of the Biden administration was introduced.

Nvidia has focused on 200 times a reduction in the cost of conclusions only in the last two years, Kress said. She also said that as AI is expanding beyond the digital world, NVIDIA infrastructure and software platforms are increasingly accepted to power the robotics and the physical development of AI. On top of that, NVIDIA’s car vertical revenue is also expected to increase.

As for CES, she noted that the NVIDIA Cosmo World Foundation platform was found there, and robotics and car companies – including Uber – are among the first to accept it.

From a geographical point of view, to the potential growth of revenue from the data center was stronger in the United States, led by the original ramp. Countries around the world are building their AI ecosystems, and demand for computing infrastructure seeking investment in 200 billion euros in the EU, and investment initiatives for 200 billion euros offer a view of the construction designed to redefine global AI infrastructure over the coming years.

Cress said that as a percentage of total revenue from the Data Center, sales of the Data Center in China remain well below the levels observed before the export control began. During no change in regulations, NVIDIA believes that Chinese shipments will remain approximately at the same level in China for data centers solutions.

“We will continue to observe export control as we serve our customers,” Cress said.

Play and ai computers

Cress noted that revenue from $ 2.5 billion games decreased by 22% in succession and 11% of the year.
A full year revenue of $ 11.4 billion increased by 9% a year, and demand remains strong throughout the holiday. But Cress said Q4 shipments were influenced by supply limitations.

“We expect a strong consistent Q1 growth with an increase in deliveries, new desktops and laptops of the GeForce RTX 50 are here, designed for gamers, creators and developers,” Cress said.

The RTX 50 Graphic cards use Blackwell architecture, fifth-generation tensor nuclei and 4th generation RT nuclei. The DLSS4 software increases the frame rate to eight times the previous generation, turning a frame depicted into three.

Car revenue was a record $ 570 million, which is 27% in succession and 103% of the year. A full year of $ 1.7 billion revenue increased by 55% a year. Strong growth is guided by the ongoing ramp in autonomous vehicles, including cars and robotics.

At CES, we announced Toyota, the largest carmaker in the world, will build its vehicles from the next generation of NVIDIA Orin, which drives NVIDIA Drive certified. Cress said NVIDIA saw higher costs for the development of engineering in the quarter, as more chips moved into production.

NVIDIA expects FYQ1 revenue to be $ 43 billion, with the consistent growth of revenue from the Bot Compute Data Center and the job network.

The next major NVIDIA event is the GTC annual conference of March 17 in San Jose, California, where Huang will deliver a major note on March 18.

Asked about the blurring between training and conclusion, Huang said that there are now “multiple scaling laws”, including the pre -training law, zooming after exercise with the help of reinforcement and scaling of test time or reasoning. These methods are at the beginning and will change over time.

“We execute every model. We are great in training. The bigger part of our calculation is actually a conclusion today. And Blackwel, with the idea of ​​referring to the models, and when you look at training, it is many times more fulfilling, “he said. “But what is truly amazing is for long thinking, reflection on test time, AI models were 10 times faster, 25 times higher bandwidth.”

He noted that today he is more enthusiastic than in CES, and noted that 1.5 million components will enter each of the trunk based on Blackwell. He said the job is not easy, but all the Blackwell partners are doing a good job. During the Blackwell ramp, gross margins will be at low percentage points in the 70s.

“At this point, we focus on speeding up our production to make sure we can provide” Blackwell Chips to customers as soon as possible, “Cress said. There is an opportunity to improve the gross margins in time until the mid-1970s later.

Huang noted that the bigger part of the software will be based on machine learning and accelerated calculations. He said the number of startup companies for AI is still quite vibrant and this agent AI for the enterprise is increasing. He noted physically AI about robotics and sovereign AI for different regions are increasing.

Blackwell Ultra is expected in the second half of the year as the next train, Huang said. He noted that the first Blacuel had a “hiccup”, which costs several months and is now completely restored.

He went to the main advantage that Nvidia has over rivals, saying that the software is “incredibly difficult” and the company builds its stack from end to end, including architecture and ecosystem that sits at the top of architecture.

Asked about geographical differences, Huang replied: “The extraction is that AI is software. This is modern software. This is an incredible modern software, and AI has passed the mainstream. AI is used in delivery services everywhere, by shopping for services everywhere. And so I think it is quite safe to say that AI has passed a mainstream, that it integrates into every app. This is already a software tool that can handle a much larger part of world GDP than at any time in history. We’re just at first. “