SiliconANGLE theCUBE · Sep 28, 2026 · 22m · 15m saved
With Gilad Shainer, Senior Vice President of Networking at Nvidia
Nvidia's last DPU, a chip that runs a data center's networking, storage and security work, secured a single 400-gigabit connection into the data center. BlueField-4, its successor, manages 7.2 terabits of secure traffic across the whole AI factory, Gilad Shainer said.
The usual picture of a data center has one network for getting in and out. Shainer said agentic AI, in which software agents chain many AI calls together, needs security inside the data center as well, on every connection between machines.
"Yes, so first agentic AI it's probably the most complex workload ever created."
Gilad Shainer, Senior Vice President of Networking at Nvidia, on SiliconANGLE theCUBE. He came back to the show, at its New York Stock Exchange studio, to explain scale-in, the networking layer Nvidia is adding for AI agents.
The full interview is covered here so you can skip it. 22 minutes of audio, 7 minutes of reading.
Here are the 10 takeaways that matter.
Key Takeaways
Nvidia now builds five kinds of network into an AI factory, and the newest, scale-in, exists to secure AI agents
BlueField-4 secures 7.2 terabits of traffic, against a single 400-gigabit pipe for the previous generation
Security for agents has to run on a separate chip the agents cannot reach, which is why it sits on the DPU and not the GPU
The monitoring runs out of band, so it does not slow the GPUs or storage
Nvidia's GPU domains grew from 8 to 72 to 576, and it now ships new AI factory generations every year
Nvidia may add a GPU to future BlueField chips to follow what agents are doing more closely
1. Agents Need a New Network
The host said Nvidia's networking had moved through several phases: NVLink to link GPUs into larger units, then InfiniBand and Spectrum-X Ethernet to connect racks and whole AI factories. He asked what agentic AI changes.
Shainer said an agent is not one question and one answer.
"There is many inference calls." — Gilad Shainer
The work runs through compute, storage, memory and networking, he said, which is why Nvidia builds Vera Rubin as a full-stack AI factory.
An AI factory needs several networks, not one
He listed them: NVLink to build a GPU unit, InfiniBand or Spectrum-X Ethernet to scale out to hundreds of thousands of GPUs, scale-across to join several AI factories when one site is not big enough, and a storage network for context, the running memory of a conversation, built with BlueField as the storage controller.
"And then recently we're also introducing a new infrastructure which is called scale in." — Gilad Shainer
