Nvidia GTC 2025: Four big announcements you need to know about
Nvidia GTC 2025, the chipmaker’s annual conference, has dominated the airwaves this week – and it’s not hard to see why. The company is the belle of the ball in global tech right now, with everyone getting in on the action.
The generative AI boom that’s saturated the industry over the last two and a half years has, by and large, been underpinned by Nvidia’s hardware. Tech giants have flocked to Nvidia’s chips to power their relentless push to achieve AI dominance.
With growing concerns over infrastructure strain, the need for great levels of compute power, and the industry pivot to agentic AI, all eyes were on Nvidia to set the stage for a big year.
The annual conference certainly didn’t disappoint. CEO Jensen Huang, donning his iconic jacket, told attendees that the industry is at the precipice of significant change.
Data center infrastructure is changing, he noted, and Nvidia wants to lead the charge toward the development of ‘AI factories’. These modern data centers are more powerful, scalable, and intelligent than their predecessors – and this evolution is being fueled by the need to run AI systems more efficiently.
“AI has made a giant leap – reasoning and agentic AI demand orders of magnitude more computing performance,” Huang said.
So what were the big announcements at GTC 2025? Here’s everything you need to know.
Nvidia GTC 2025: Blackwell Ultra AI Factory Platform
Among one of the first major announcements at GTC 2025, Nvidia unveiled the launch of its Nvidia Blackwell Ultra AI Factory Platform.
The launch of the platform, Nvidia said, paves the way for the ‘age of AI reasoning’, and will help users boost training and test-time scaling inference capabilities. This essentially means enterprises can apply more compute during inference to improve accuracy.
As the name suggests, the platform is built on the Blackwell architecture, and includes the Nvidia GB300 NVL72 rack-scale solution, which the company said can deliver 1.5x improved AI performance compared to its predecessor.
This launch comes at a pivotal time in the evolution of the AI space, with enterprises now flocking to agentic AI, the company said. Accuracy is the name of the game amidst this shift, and by giving enterprises the ability to rapidly scale inference capabilities, Nvidia is betting this could be a game changer.
“We designed Blackwell Ultra for this moment,” Huang said. “It’s a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference.”
Nvidia launches a family of open reasoning models
Agentic AI was once again given a major boost at GTC with the launch of Nvidia’s Llama Nemotron open reasoning models. This family of models is designed specifically to support developers to create advanced AI agents.
Built on Meta’s Llama models, this offering will provide users with on-demand AI reasoning capabilities, and the company said it advanced the family during post-training to improve areas such as multi-step math, coding, reasoning, and complex decision making.
“This refinement process boosts accuracy of the models by up to 20% compared with the base model and optimizes inference speed by 5x compared with other leading open reasoning models,” Nvidia said in a statement.
These improvements mean the models can handle more complex reasoning tasks – making them the perfect foundation for enterprise AI agents – while reducing operational costs through improved efficiency.
Several industry big hitters are already working with the Llama Nemotron models, according to Nvidia. Microsoft, for example, has integrated the reasoning models within Microsoft Azure AI Foundry.
SAP is also tapping Llama Nemotron models to advance SAP Business AI solutions and Joule, the company’s flagship copilot tool.
Isaac GR00T N1
Advancements in robotics were a key talking point – and source of excitement – at GTC 2025, and Nvidia added more fuel to the fire with the launch of Isaac GR00T N1, the world’s first open source humanoid robot foundation model.
This is a family of fully customizable models, pre-trained by Nvidia, that will be released to robotics developers globally. The aim here is to accelerate the development of ‘generalist robots’ that can perform tasks akin to their human counterparts.
The potential use-cases are endless, and Huang went so far as to suggest that AI-powered robotics will be one of the big trends in the industry in years to come.
“The age of generalist robotics is here,” he said. “With Nvidia Isaac GR00T N1 and new data-generation and robot-learning frameworks, robotics developers everywhere will open the next frontier in the age of AI.”
GR00T N1 can “easily generalize across common tasks”, the company noted. This includes grasping, moving objects with one or both arms, and transferring items from one arm to another.
In his keynote, Huang showed off a humanoid robot developed by 1X, which performed domestic tasks such as tidying up.
Vera Rubin Ultra
Nvidia gave attendees a tantalizing glimpse into the future of its GPU roadmap this week with the Vera Rubin AI superchip. The successor to the company’s Grace Blackwell chip, Vera Rubin also uses a combination of a CPU and GPU.
While ‘Grace’ represents the CPU and ‘Blackwell’ the GPU, in this instance ‘Vera’ and ‘Rubin’ match accordingly.
Nvidia said Vera will boast 4.2 times the memory and 2.4 times the memory bandwidth compared to Grace, while Rubin will offer 288GB of high-bandwidth memory 4 (HBM4) compared to the HBM3e included in the Blackwell Ultra.
Exciting developments, no doubt, but Nvidia went one step further, unveiling plans for its Vera Rubin Ultra chip. These will provide even better performance stats upon launch, but we could be waiting a while – they’re expected to ship in the second half of 2027.
The big talking point here was the power consumption, with the company revealing a rack of Vera Rubin Ultra chips will consume up to 600kW each.
MORE FROM ITPRO
Source link
