Whenever Nvidia hits the road for its GTC shows, the tech industry watches closely to see what the world's most valuable company has cooked up.

This time out didn’t disappoint, with a revolutionary way to interconnect quantum computers with classical systems and a bumper $1 billion team-up with Nokia around AI-augmented radio access networks (RAN).

But chief among its biggest unveils was a product for which its seeds were sown back in 2021: BlueField-4, a data processing unit – or DPU – that’s set to quietly help shore up some of the largest AI factories.

Itay Ozery, Nvidia’s director of product marketing for networking, described BlueField-4 as the “hero announcement” of this month’s GTC, likening it to becoming the “processor for running the operation system of the AI factory.”

Half a decade in the making

BlueField-4 was first teased at GTC 2021, a time when AI was not the world-encapsulating concept it is today. Instead, workloads were general-purpose, cloud was king, and InfiniBand was winning the battle for scale-up.

But fast forward to 2025, and it’s all about AI. Despite not commanding the limelight in the way Nvidia’s GPUs do, its BlueField series would also receive the AI overhaul, with CEO Jensen Huang proclaiming the DPU as “purpose-built for the gigascale AI era.”

“What we've done since 2021 is we realized that, hey, we need to optimize the AI compute fabric so it can support the scaling laws of AI with the Hopper and Blackwell generation needing to scale to tens-of-thousands of GPUs and even hundreds-of-thousands of GPUs,” Ozery explained.

Among the biggest changes that helped define BlueField and its architectural importance is Nvidia’s Ethernet-based Spectrum-X networking fabric.

Ozery outlined that the third generation of BlueField helped define Spectrum-X, leveraging its related IP and architecture to optimize the fabric to support AI workloads.

That innovation, based on its prior generation, helped create an offering built on a protocol the tech giant has typically not been known for, which has since gone on to form the basis of major customer wins, most notably with the likes of Meta and Oracle.

Now it’s the turn of BlueField to get its moment in the spotlight, with the newly launched fourth generation offering some six-times the compute power compared to the prior generation.

Powering that performance improvement is the addition of a Grace CPU that’s packed with ever-more powerful cores. But in addition to yet another powerful chip is a second new innovation on display at GTC: the ConnectX-9 SuperNIC.

“We’re combining the best of both worlds,” Ozery said. “The CPU for the compute and the ConnectX-9 for the network IO. We’re packaging it up. And In a way, we’ve continued to innovate while bringing innovation to the AI compute fabric.”

Software-defined? No, software-led

When it comes to the launch date of BlueField-4, then, when compared to Nvidia’s newfound GPU product cadence, or its so-called “annual rhythm,” BlueField is a little more spread out.

Case in point: when BlueField-4 was first hinted at in 2021, it was slated for 2024. Now it’s set for 2026 as part of Nvidia’s Vera Rubin platform, with its eventual successor, the BlueField-5, hinted at for 2028.

Ozery told SDxCentral that compared to Nvidia’s GPU roadmap, he didn’t feel the cadence of its BlueField hardware was as important.

“You will see it right [at GTC] today. We have partners that will launch new solutions, new applications, on top of BlueField-3. And the reason why is because they believe in this DOCA architecture. They can build once, and it will run seamlessly and faster.”

DOCA, or the data center infrastructure-on-a-chip architecture. It’s a software framework to manage Nvidia’s arsenal of DPUs and network cards, providing operators with APIs and tools to accelerate and isolate workloads – similar to how CUDA works for GPUs.

According to Ozery, Nvidia’s wider efforts around building out its software stack and making sure CUDA is front and center as the underlying component augmenting its hardware offerings have helped drive innovation for its networking offerings via DOCA.

“We're basically taking all of the same strategy for AI compute, GPU compute, and applying it to infrastructure processing with BlueField and the DOCA architecture.”

Technical details: Bandwidth bits

For the technically minded readers, SDxCentral got an inside track on the actual topology of BlueField-4.

The Nvidia director revealed the new DPU will offer 800Gb/s total bandwidth, adding: “In most cases, you'll see two 400Gb/s ($00G) links, bi-directional.”

In addition, according to Ozery, most deployments will run using Ethernet at the front end, handling what the north-south traffic – the data flows between servers and storage systems, internet access for inference workloads, and the security services running directly on the BlueField-4 itself.

For GPU-to-GPU communications – the "east-west" traffic that powers training and large-scale inference – Nvidia is packaging multiple ConnectX-9 chips together to deliver a whopping 1.6Tb/s per GPU starting with next year's Rubin platform.

BlueField-4's focus on 800Gb/s for cloud and storage networks reflects what Ozery sees as sufficient headroom for the Rubin generation.

“Just by moving from 400 to 800, we believe this is meeting the requirements for everything around the Rubin architecture,” he said, though hinted that faster connectivity speeds would follow in future generations.

On the storage front, BlueField-4 builds on capabilities introduced in earlier generations, including NVMe over Fabrics (NVMe-oF) and NVMe/TCP support.

Partners like VAST Data are already running their entire storage stack directly on BlueField hardware, a pattern Ozery expects to accelerate with the fourth generation as AI workloads demand ever-faster data access for training, inference, and retrieval workflows.

Deployments of BlueField by operators like Vast have further validated Nvidia's strategic patience, according to Ozery. From hyperscalers to cybersecurity vendors, OEMs to storage partners, he opined that the industry is leaning into BlueField's eventual evolution.

“They're telling us, hey, we want to leverage BlueField in order to utilize our GPU compute to the fullest potential,” Ozery said. “We have BlueField that is running all of the infrastructure workloads, your storage, data movement, cybersecurity – everything you will run on BlueField, and they can purpose GPU compute, Blackwell, Rubin, for running AI applications. This is why they're building these massive AI factories.”