Jensen Huang stated that last year the four major cloud service providers in the USA purchased 1.3 million Hopper architecture chips, and this year they have bought 3.6 million Blackwell chips. It is expected that by 2028, capital expenditure in datacenters will exceed 1 trillion USD; General Motors will use NVIDIA technology to help develop autonomous vehicles, and NVIDIA has launched the automotive safety AI solution Halos; NVIDIA will cooperate with telecom companies like T-Mobile US to develop AI networks for 6G; Blackwell architecture has gone into full production, and customer demand is Incredible; the "AI factory operating system" Dynamo has been launched, with Blackwell NVLink72 chips delivering 40 times the inference performance of Hopper.
On local time March 18th, Tuesday, $NVIDIA (NVDA.US)$ CEO Jensen Huang delivered a keynote speech at NVIDIA's AI event GTC 2025 held in San Jose, California.
Jensen Huang stated that last year's GTC conference was hailed as the Woodstock music festival of the AI field, while this year’s GTC is referred to as the American version of the AI industry's Spring Festival Gala, the "Super Bowl"; the only difference between these two descriptions is that in the "Super Bowl", everyone is a winner.
He first introduced the development history of AI research and development, from the initial Perception AI to the current Generative AI. He expects that we will enter the era of Agentic AI, followed by the era of Physical AI, which is the era of robots.
Jensen Huang said that we are currently understanding how to scale AI, and in the future, attention must be paid to training and scaling AI models that are built. He introduced the evolution of the scaling law, from pre-training scaling, post-training scaling, to test-time scaling, known as "long thinking."
Jensen Huang believes that the industry misjudged the demand for computing last year. He said:
The demand for computing, or the expansion law of AI, is more resilient, and in fact, the speed is increasing at a hyper-accelerated rate.
The four major cloud computing service providers in the USA have purchased 1.3 million Blackwell chips this year.
Jensen Huang stated that the amount of computation required for inference has significantly increased compared to before, while the data and human training available are limited. In the future, there will be a transition from human-written Software to Software run by AI models.
Jensen Huang introduced that the growth of AI computing-related infrastructure is at a turning point.
He revealed that in 2024, the top four cloud service providers in the USA, known as hyperscalers, purchased 1.3 million NVIDIA Hopper architecture chips, and in 2025, they purchased 3.6 million Blackwell architecture chips.
Jensen Huang expects that by 2028, the capital expenditure for building Datacenters will exceed 1 trillion dollars.
Jensen Huang showcased NVIDIA's simplified acceleration platform processing and the CUDA-X library adopted in fields such as data and AI, stating that AI acceleration services can be applied across various Industries and that this is only a small part of the libraries for achieving accelerated computing.
Jensen Huang predicts that every company will have two factories in the future, one for producing products and the other for AI mathematics. Huang claimed that AI will enter all Industries.
NVIDIA announced its next-generation super chip, Vera Rubin.
NVIDIA CEO Jensen Huang showcased the next-generation Vera Rubin AI super chip and Blackwell Ultra at the GTC conference. He stated that the transition to the Blackwell Ultra chip will occur in the second half of this year, with Vera Rubin replacing the Blackwell Ultra chip starting in the second half of 2026.
Vera Rubin is similar to Grace Blackwell, integrating both CPU and GPU. In Grace Blackwell, Grace is the CPU while Blackwell is the GPU; in Vera Rubin, Vera is the CPU and Rubin is the GPU.
NVIDIA stated that the memory of Vera CPU is 4.2 times that of Grace, and the memory bandwidth is 2.4 times that of Grace. With Vera's 88 CPU cores combined, NVIDIA claims that the overall performance of this chip will be twice that of the previous generation product. The Rubin GPU will be equipped with 288GB of HBM4.
Furthermore, NVIDIA announced the next generation of chips after Vera Rubin, named Vera Rubin Ultra. Set to be released in the second half of 2027, Vera Rubin Ultra will combine the Vera CPU with the Rubin Ultra chip. Each Rubin processor consists of two GPUs in a single chip, while Rubin Ultra consists of four GPUs.
Subsequently, Jensen Huang announced in a roadmap PPT that the next generation after Rubin is named Feynman, after the renowned physicist Richard Feynman. According to NVIDIA's roadmap, the Feynman architecture will debut in 2028.
Collaborating with General Motors to develop autonomous vehicles and working with companies like T-Mobile US to develop AI networks for 6G.
Jensen Huang announced that NVIDIA will expand cooperation with $General Motors (GM.US)$ General Motors will utilize NVIDIA's technology to help develop autonomous vehicles and train AI manufacturing models using NVIDIA's technology.
NVIDIA launched an AI solution focused on automotive safety, named NVIDIA Halos. Jensen Huang said, "I believe we are the first company in the world to conduct safety assessments on every line of code."
Jensen Huang also announced that NVIDIA will be collaborating with $Cisco (CSCO.US)$ and $T-Mobile US (TMUS.US)$ Collaborate with companies to research and develop AI-native networks for next-generation wireless network 6G.
The Blackwell architecture has entered full production, launching the "AI Factory Operating System" Dynamo.
Speaking about Datacenters, Huang Renxun stated that the chips of the Blackwell architecture are now in full production, "customer demand is Incredible."
He once again showcased the super chip Grace Blackwell NVLink 72 that he demonstrated at CES this January. It integrates 72 Blackwell GPUs on a single wafer and has 18 NVLink Switches, achieving a computational performance of 1.4 EFLOPS on 4-bit floating point FP4.
NVIDIA has launched what is called the 'AI Factory Operating System' Dynamo. It is a 'distributed inference service library'. It is essentially an open-source solution to address the issue of not being able to provide enough tokens needed by users.$Microsoft (MSFT.US)$Perplexity is one of the first partners of Dynamo.
Huang Renxun demonstrated how the Blackwell architecture surpasses the Hopper supercomputer. Equipped with the Grace Blackwell NVLink72 chip and Dynamo, the Blackwell architecture can enhance performance by 25 times compared to the Hopper architecture. "In inference models, the performance of Blackwell is 40 times that of Hopper."
Huang Renxun jokingly said that the Hopper is sufficient for some tasks, but with the advent of Blackwell, "I am the head of revenue destruction." With the support of the latest technologies like Blackwell, manufacturers building AI factories will find that "the more you buy, the more you save."
Introduced the world's first open-source humanoid robot functional model.
NVIDIA launched the world's first open-source humanoid robot functional model, Isaac GR00T N1, and introduced Simulation Frameworks to accelerate robot development.
The robot Blue, developed in collaboration between NVIDIA, Alphabet-C, and Disney, was unveiled.
The last segment of the keynote speech at NVIDIA GTC 2025 focused on robots, with Jensen Huang announcing that NVIDIA, $Alphabet-C (GOOG.US)$ DeepMind, $Disney (DIS.US)$ are collaborating to develop a robotic platform named Newton.
NVIDIA first showcased a demonstration animation, then a real Siasun Robot&Automation appeared on stage and started walking towards Huang Renxun. This robot moved like the robots in 'Star Wars', making adorable sounds and walking naturally. Huang Renxun named this robot Blue, which is equipped with NVIDIA's latest GR00T N1 general-purpose model.
Editor/rice
