Nvidia’s DGX AI Systems Are Faster and Smarter Than Ever 

A DGX SuperPOD. (Source: Nvidia)

At its GTC event in San Jose today, Nvidia unveiled updates to its AI infrastructure portfolio, including its next-generation datacenter GPU, the NVIDIA Blackwell Ultra.

Expanding on the Blackwell architecture introduced last year, Nvidia is integrating its new Blackwell Ultra GPUs into two DGX systems: the NVIDIA DGX GB300 and the NVIDIA DGX B300.

The DGX GB300 system, designed with a rack-scale, liquid-cooled architecture, is powered by the Grace Blackwell Ultra Superchip, which combines 36 NVIDIA Grace CPUs and 72 NVIDIA Blackwell Ultra GPUs. The DGX B300 system, an air-cooled alternative, leverages the NVIDIA B300 NVL16 architecture.

For large-scale deployments, customers can combine multiple DGX systems into a DGX SuperPOD—a pre-configured AI infrastructure that integrates compute, networking, storage, and software. Effectively, a DGX SuperPOD functions as an AI supercomputer, or as Nvidia CEO Jensen Huang calls it, an AI factory.

Ahead of GTC, AIwire spoke with Nvidia Vice President of DGX Systems, Charlie Boyle, about the new DGX systems and the 70x speedup in AI inference and reasoning they deliver over the previous Hopper-based generation.

Scaling AI Beyond Training

As AI adoption accelerates, companies are moving beyond just training models and need to deploy them at scale for real-time applications. Inference, and more specifically AI reasoning, has become a critical workload, requiring systems that can handle the growing demand for speed and efficiency.

“The AI factory does a number of things. It can do training. It can do inference. A lot of the talk in the past year has been about reasoning, which is a form of inference,” Boyle said, noting that the DGX systems have been known as great systems for training and post-training, but customer workloads are now pivoting toward inference and reasoning.

Nvidia's new Blackwell Ultra GPU. (Source: Nvidia)

Nvidia achieved the 70x speedup in AI reasoning through a combination of hardware and networking advancements in the DGX GB300 system, Boyle says. At the core of this improvement is the Blackwell Ultra GPU, which introduces faster FP4 precision math and expanded memory capacity to significantly accelerate inference workloads.

Complementing the upgraded compute power is Nvidia’s newest ConnectX-8 networking technology, which enables 800-gigabit-per-second connectivity across nodes within the rack for ultra-fast data transfer between GPUs. Boyle notes that as AI models scale to support thousands or even millions of users, efficient networking becomes critical, and ConnectX-8 allows thousands of DGX racks to interconnect to form large-scale AI factories. Additionally, ConnectX-8 supports both Ethernet and InfiniBand, giving customers the flexibility to optimize their network architecture for their specific workloads.

The gains in reasoning also come with improvements in energy efficiency, stemming from a combination of increased compute performance and a more efficient power subsystem design. Instead of using individual power supplies for each node, the GB300 and B300 DGX systems incorporate a rack-wide power bus bar and centralized power shelf technology. This approach reduces power conversion losses, optimizes energy distribution, and eliminates the inefficiencies associated with over-provisioning, the company says.

Traditionally, datacenters must reserve extra power capacity to accommodate peak loads, leading to wasted “stranded power.” Nvidia’s new power management system smooths out these fluctuations, Boyle says, allowing datacenters to deploy more GPUs without unnecessary energy overhead.

“By eliminating [stranded power], we allow customers to deploy more systems and have less cost overall, because you're not paying for power that you're never using. You're getting the most out of your datacenter,” Boyle says.

This efficiency gain is critical for large-scale AI deployments, where infrastructure must scale to support hundreds of thousands of GPUs without excessive power consumption. By maximizing the use of available energy, Nvidia says it is enabling higher AI throughput per megawatt, reducing operational costs while improving sustainability for large-scale AI infrastructure.

How Customer Feedback Shapes DGX

Customer feedback has played a critical role in shaping Nvidia’s latest DGX advancements, particularly as AI infrastructure scales beyond research labs and into enterprise environments.

Boyle, who joined Nvidia in 2016 ahead of the first DGX system launch, has spent years cultivating relationships with DGX customers worldwide.

“I have some customers that have every generation system, all the way back to the DGX-1, to our latest systems. And we listen to our customers,” he says.

Over the years, thousands of DGX customers, many of whom have deployed multiple generations of the system, have provided insights that directly influenced Nvidia’s design decisions. One of the biggest takeaways has been a simple yet powerful request: Just getting the work done.

“At the end of the day, if you're an AI researcher, you really don't care about infrastructure. You just want it to work,” Boyle says.

Originally built for AI researchers, DGX systems are now widely used by IT teams and datacenter operators, making ease of deployment and operational efficiency more critical than ever. To address this, Nvidia has introduced Mission Control, a software stack designed to streamline AI infrastructure management from installation to daily operations. It automates cluster bring-up, job scheduling, failure recovery, and resource optimization, ensuring that AI users can focus on their workloads rather than infrastructure issues.

The need for resilient, self-managing AI systems becomes even more evident as clusters grow in complexity. Customers have expressed frustration over lost productivity due to unexpected failures. If a job fails overnight, hours of compute time are wasted. Mission Control solves this problem by automating job restarts, checkpointing progress, and optimizing system efficiency in real-time. Built on years of internal Nvidia expertise, the platform delivers the same operational intelligence that powers Nvidia’s own infrastructure, Boyle says, ensuring customers benefit from the company’s deep experience in managing large-scale AI clusters.

“This is all based on tools, technologies and techniques that we've developed over the past decade inside of Nvidia. That's one of the core bases of the DGX platform, the great work that all of our engineers do internally. We package that up and make that available to customers,” Boyle says.

Inside the DGX User Group at GTC

Nvidia's focus on DGX customers extends to GTC, where Boyle will connect with them in person. “We run a great user group event every year at GTC. I'm seeing them all bright and early Wednesday morning here,” Boyle says.

The DGX User Group at GTC is an exclusive, sold-out gathering for DGX customers, offering a deep dive into new technologies, real-world deployments, and future AI infrastructure plans. Unlike the broader announcements from Jensen Huang’s keynote, this session is a highly technical, hands-on forum where users can explore the finer details of DGX advancements.

Each year, the session features customer presentations where AI practitioners share their experiences from deploying DGX-powered AI infrastructure. Attendees gain insights into real-world AI factory operations, AI reasoning, and software optimizations directly from their peers. The event also provides an interactive space for customers to ask Nvidia’s product managers and engineers detailed technical questions, ensuring they can maximize performance and efficiency in their own deployments.

Beyond product discussions, the DGX User Group is also about community building. With customers worldwide running identical DGX hardware and software, the event fosters a unique knowledge-sharing environment, where attendees can exchange best practices, troubleshooting tips, and scaling strategies. Nvidia further enhances this by analyzing customer usage data, sharing trends and benchmarks to help users understand where they stand relative to their peers.

The session isn’t just about immediate problem-solving but is also about preparing for the future. Boyle and his team provide guidance on Nvidia’s AI infrastructure roadmap, helping customers plan for upcoming advancements. For many attendees, this closed-door session is a highlight of GTC, bringing together some of the most advanced AI users in the world for a deep technical exchange.

Instant AI Factory Brings AI Deployment on Demand

As AI adoption accelerates, companies are moving beyond isolated training experiments and into large-scale production deployments. Many organizations start with internal AI applications, testing them with a small group of users. But once these tools prove their value, demand skyrockets, sometimes growing from dozens to thousands of users almost overnight. This rapid scaling presents a new challenge: how to deploy AI efficiently while ensuring infrastructure keeps pace with demand.

Nvidia’s DGX platform is designed for this kind of scalability, with a consistent software stack that has remained compatible across nine generations of hardware. However, one of the biggest challenges in scaling AI isn’t just compute power—it’s datacenter capacity.

To address this, Nvidia has unveiled the NVIDIA Instant AI Factory, a managed service featuring the Blackwell Ultra-powered DGX SuperPOD. Equinix will be the first provider to offer DGX GB300 and DGX B300 systems in its preconfigured liquid- or air-cooled AI-ready datacenters, spanning 45 global markets, according to Nvidia.

These facilities are pre-plumbed for liquid cooling and optimized for DGX deployments, allowing customers to scale quickly without having to navigate complex datacenter engineering requirements. Instead of waiting months or even years to secure infrastructure, companies can now spin up AI capacity in days by simply specifying how many racks they need and where they need them.

Nvidia says DGX SuperPOD systems with DGX GB300 or DGX B300 will be available from partners later this year, with NVIDIA Instant AI Factory also expected to launch later this year.

As GTC unfolds, Boyle is eager to see how Nvidia’s partners and customers are applying the company's innovations in the real world, which is something he looks forward to the most.

“We've got a tremendous amount of customer speakers at GTC just sharing their experiences of what they've done, and it's always incredible,” Boyle shares. “We've got some of the top researchers working internally at Nvidia, and I get to see all the great work they're doing. But when I hear customer stories about how something really changed their business or changed the way they work, and how easy or hard something was ... just hearing all those stories always inspires me.”