AI infrastructure is undergoing a fundamental shift as the unit of deployment moves from the server rack to the entire data center, making open co-design networking a first-class design priority rather than an afterthought.

That reordering was on full display this week as Advanced Micro Devices Inc. used its Advancing AI 2026 event to argue that its Helios rack-scale platform can’t succeed without radical openness in networking. The company detailed how monthly token consumption has surged 158 times in just two years, a trajectory that is pushing network throughput, reliability and programmability to the center of AI system design, according to Soni Jiandani (pictured, right), senior vice president and general manager at AMD.

“The role of the network is only increasing because if you have to deliver these systems at scale with performance, with high availability, and this type of token consumption of 158X in two years, the network is foundational to deliver a balanced approach,” Jiandani said. “With the evolution towards these large frontier models and agentic AI, the network couldn’t be more foundational.”

Jiandani and Omar Baldonado (left), senior director of data center and AI networking at Meta Platforms Inc., spoke with theCUBE’s Dave Vellante at the AMD Advancing AI event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how open co-design networking underpins AMD’s Helios rack-scale strategy, the role of Ethernet in AI back-end deployments and how agentic workloads are intensifying networking demands. (* Disclosure below.)

Open co-design networking facilitates shared AI infrastructure

That co-design networking extends deep into the fabric itself, where AMD and Meta jointly built the scale-up interconnect that binds Helios’ GPUs into a single resource pool. The work reflects a broader industry pivot, as Ethernet has now surpassed InfiniBand in AI back-end network deployments.

“We partnered very closely with Meta to make sure that the Ultra Accelerator Link, or UALink, over Ethernet … becomes the fabric to bring together all the GPUs,” Jiandani said. “We have 50% more memory than our competitor. How do we bring high availability attributes? Because now the failure domain is dozens of GPUs acting as one with a lot of memory.”

The openness embedded in that partnership traces back to Meta’s early role in founding the Open Compute Project’s networking effort, an approach that has directly shaped Helios, Baldonado noted. Networking demands are compounding as inference and agentic workloads pull in CPU access, storage and distributed processing simultaneously.

“Networking needs the openness, because it’s about plugging things together,” Baldonado said. “So many different technologies across the stack have to interoperate and work together.”

The shift toward agentic AI is intensifying that complexity further, as context-heavy workloads require vast key-value, or KV, cache data shared across servers and storage. AMD’s data processing units, or DPUs, are being positioned to absorb that burden, freeing GPUs and CPUs for higher-value work, Jiandani noted.

“No one company can figure this out alone,” Jiandani said. “Having the opportunity to co-design, co-develop, think about where the next generation of workloads will land gives you the opportunity not just to bring the best of assets that we have, but also to bring the best of assets with our customers along.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the AMD Advancing AI event:

(* Disclosure: TheCUBE is a paid media partner for the AMD Advancing AI event. Neither AMD, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE