Nvidia's AI Super Legion is here! Hwang In-hoon says: Demand for AI computing power will skyrocket 100 times

source元宇宙之心MetaverseHub·元宇宙之心MetaverseHub·00:37 编辑
Nvidia's AI Super Legion is here! Hwang In-hoon says: Demand for AI computing power will skyrocket 100 times

Nvidia founder and CEO Huang Renxun told thousands of viewers at the company's annual AI developer conference on Tuesday that artificial intelligence is at a “critical inflection point.”


At GTC 2025, known as the “Super Bowl of the AI World,” Hwang In-hoon's keynote address focused on Nvidia's latest breakthroughs in the AI field and shared his predictions for the industry's development in the next few years.


Image


He pointed out,Demand for GPUs from the world's top four cloud service providers has surged, and Nvidia's data center infrastructure revenue is expected to exceed $1 trillion in 2028。

Hwang In-hoon said that in order to promote the development of artificial intelligence, especially moving towards AI agents (AI agents) and inference AI models (Inference AI models), strong computing power is needed, even far beyond the current level.


This trend towards AI agents and inference AI means “a significant increase in the amount of computation required to train and infer these models.”


In contrast, traditional big language models (LLM) have lower computational requirements and can generate answers instantly. The inference model, on the other hand, requires multiple rounds of internal reasoning before an answer, so the computational requirements are much higher and the response time is longer.

“In order to maintain the model's responsiveness so that users don't lose patience due to waiting, we now need to increase the calculation speed by 10 times,” said Hwang In-hoon. “The overall computing demand can easily grow to 100 times.”

01.
Nvidia GPUs are still essential

Hwang In-hoon's remarks are meant to emphasize that the AI industry still needs a large number of Nvidia GPUs.


However, in January of this year, AI startup DeepSeek revealed that they only used 2,000 slower Nvidia H800 chips to train high-performance basic AI models, while companies such as OpenAI usually require tens of thousands or more GPUs.


This news once raised concerns in the market, causing Nvidia's stock price to plummet, and the market value evaporated by nearly 600 billion US dollars within a day. Wall Street investors feared that GPU demand was overestimated.


However, Hwang In-hoon believes that the future development of AI agents and inference AI will bring greater demand. He predicted,In the future, the world's 1 billion knowledge workers will have 10 billion AI agents working together。


Strong growth in market demand is already being traced.


Hwang In-hoon revealed that in the year demand for Nvidia Hopper GPUs was at its peak, the company delivered 1.3 million chips to the four major cloud computing companies AWS, Microsoft, Google, and Oracle. In its first year of launch, the latest Blackwell-architecture GPUs have already shipped 3.6 million units.


Image


Also, during the presentation, Hwang In-hoon showed an AI model duel — Meta's LLAMA open source model and DeepSeek's R1 inference model. The user asked the two models a wedding seating arrangement question: at the 7-seat table, make sure that the bride and groom's parents are not next to each other, and that other restrictions are met.


Llama quickly gave an answer, generating 439 tokens (each token is about 0.75 words), but the answers were wrong. Although R1 answered correctly, the calculation time was longer and 8,559 tokens were generated. Since users are charged per token, the calculation cost is also higher.


Hwang In-hoon said,Although optimization technology can improve AI computing efficiency and thereby reduce the consumption of computing resources, overall demand will continue to grow。


AI computational efficiency is also a problem that many startups are overcoming. For example, Inception Labs, co-founded by Stanford University, UCLA, and Cornell University professors, is developing parallel computing technology to change the AI token generation process from traditional one-by-one generation to parallel generation, thereby reducing GPU computation time.

02.
Highlights: Blackwell Ultra and Rubin AI chips

During the much-anticipated launch, Hwang In-hoon revealed more details about Nvidia's next-generation GPU architecture:

  • Blackwell Ultra: Scheduled to launch in the second half of 2025;
  • Vera Rubin (RubinAI chip, named after famous astronomer Vera Rubin): Expected to be released by the end of 2026;
  • Rubin Ultra: Expected to debut in 2027.


In his speech, which lasted more than two hours, Hwang In-hoon reviewed “extraordinary progress” in the field of AI.


Image


He said that in the past ten years, AI has evolved from initial perception and computer vision to generative AI, and now it is moving towards intelligent AI (Agentic AI) with inference capabilities.This means that AI can not only understand and generate content, but also has the ability to make autonomous reasoning and intelligent decisions。


The release of GTC 2025 marks Nvidia's entry into a new stage in the field of AI computing, and also indicates that the future development of AI will enter a higher dimension.


03.
Isaac GR00TN1: Helping the development of humanoid robots

To promote robot AI training, Huang Renxun released Isaac GR00TN1, an open source basic model designed specifically for the development of humanoid robots. This model will be combined with the upgraded Cosmos AI to generate simulated training data to help robots learn more efficiently.

“It's really exciting. “Providing an open source platform means more people can learn reinforcement learning techniques,” said Benjamin Lee, a professor of electronics and systems engineering at the University of Pennsylvania. “Not only industry giants, but academic researchers can also experiment with this synthetic data.”

Hwang In-hoon also mentioned,Cosmos series AI models can generate high-fidelity videos at low cost for training robots and automation systems. He first introduced this series of models at the CES conference at the beginning of this year.

Image


This open source model can be seamlessly connected to Nvidia's Omniverse physical simulation tool, making AI training videos more realistic while drastically reducing traditional data acquisition costs. In contrast, traditional methods such as having cars collect data on real roads, or having humans train robots repeatedly, are extremely expensive and time-consuming.

04.
Nvidia signs partnership with GM

In addition to AI computing requirements, Nvidia also announced a partnership with General Motors (GM) to jointly develop AI customized systems for automotive manufacturing, factory management, and robotics applications.

GM's factories and robots will be optimized with Nvidia Omniverse and Cosmos AI base models. These technologies can be used to create a “digital twin” (Digital Twin) of an assembly line, reducing downtime through virtual testing and production simulations.

In addition, the two sides also plan to train existing robots to complete tasks such as material handling, transportation, and precision welding more efficiently.

On the automotive side, GM plans to use Nvidia's Drive AGX system to improve advanced driver assistance systems (ADAS) and safety performance. Previously, GM had already used Nvidia GPUs for AI training, and this collaboration has further expanded into the design and operation of automotive plants.
05.
Other Nvidia collaboration projects

Nvidia also announced a partnership with Google and its parent company Alphabet to accelerate the application of AI in robotics, healthcare, manufacturing, and energy. Engineers from both sides are working together to develop robots with handling capabilities, reshape drug development processes, and optimize energy networks.

Meanwhile, Nvidia is also partnering with GE Healthcare to develop autonomous X-ray technology and ultrasound applications using its latest Isaac for Healthcare medical device simulation platform. The platform integrates pre-trained models and physics-based simulations to train and verify autonomous imaging systems before actual deployment, improving accuracy and reliability.

The goal of this collaboration is to expand global access to medical imaging technology. Currently, about two-thirds of the world's population still has no access to advanced medical imaging services, and by introducing robotics, this collaboration hopes to benefit more patients.
06.
Desktop-class supercomputer launched

Nvidia also unveiled the Nvidia DGX series of desktop supercomputers, using its Grace Blackwell platform to provide powerful AI computing power.

Image


The series includes two models:


  • DGX Spark: Previously unveiled as “Project Digits,” it is a handheld supercomputer that can provide up to 1000 trillion operations (TOPS), and is specially designed for AI fine-tuning and inference.
  • DGX Station: A desktop-level supercomputer with datacenter-level computing performance, suitable for high-intensity AI computing tasks.

According to Nvidia, AI developers, researchers, data scientists, and students can use these devices to prototype, fine tune, and reason models locally or in the cloud. In addition, well-known manufacturers such as Asus, Dell, HP, and Lenovo will be responsible for the production of these computing devices.

Also,In the field of quantum computing, Nvidia is also speeding up its deployment, announced the establishment of the Nvidia Accelerated Quantum Research Center (Nvidia Accelerated Quantum Research Center) in Boston. The center will combine quantum hardware and AI supercomputers to promote the development of quantum computing technology, and is scheduled to be officially put into operation in 2025.

At the end of his speech, Hwang In-hoon released Newton, an open source physics engine designed specifically for robot simulation. This engine was jointly developed by Nvidia, Google DeepMind, and Disney Research to improve the robot's training effectiveness in a simulated environment and make it closer to the physical laws of the real world.

Afterwards, a special “guest” was welcomed to the scene — Blue, a small square robot. It jumped out of a secret door on the stage floor, made an electronic voice in response to Hwang In-hoon, acted according to instructions, and eventually stopped next to him.

Hwang In-hoon announced with a smile: “The era of general-purpose robots has arrived.”


Original Link
#元宇宙之心MetaverseHub
说明: All Bitpush articles reflect the author's views only and do not constitute investment advice.

Related

Loading...