Interview: AT&T completed a controlled introduction phase of its cloud RAN deployment, with roughly 25 sites running live traffic on Intel’s Sapphire Rapids processors as it plans to aggressively scale deployment when Intel’s next-generation Granite platform arrives over the coming months.

The Sapphire deployment is currently handling the TDD layers of the network, which represent the largest share of traffic, but not yet the full stack.

Rob Soni (pictured), VP of the operator’s RAN architecture, told Mobile World Live (MWL) Granite, which is already in AT&T’s labs, will enable a full stack covering FDD, TDD, NR and LTE from a single server design.

“This is all coming in phased pieces through the fall and early next year,” he said. “We plan to scale more aggressively on Granite than we did on Sapphire.”

He said AT&T expects to move from a few hundred Sapphire sites to thousands on Granite through 2026 and into 2028.

Performance on the Sapphire sites is matching traditional RAN on call drop rates and throughput, Soni said, and the team is beginning to utilise the platform’s promise of faster feature delivery.

The leading example is AI-driven link adaptation, developed with Ericsson, which replaces rules-based scheduling algorithms with a deep neural network.

Trials are showing roughly 10% gains in spectrum efficiency and 15% improvements in individual user throughput, with further benefits expected as AT&T tunes the global Ericsson model to its own network morphology.

“We’ve done a lot of trailing with the kind of global model of Ericsson,” Soni explained. “We’re now in the process of tuning the model to the AT&T morphology, the AT&T usage. That will change the nature, and we expect there to be more gain.”

Halfway point

In cellular networks, Soni stated a base station must determine on every transmission frame which users to serve, what modulation and coding scheme to use and how to balance competing demands across the spectrum.

It is a computationally intense optimisation problem, and the traditional approach has been to encode engineering judgment into static rule sets. The AI model Ericsson developed learns from network behaviour and adapts more fluidly to changing conditions.

Link adaptation is a critical algorithm for determining the overall multiple user performance in a cellular network,” he stated.

Rival T-Mobile US stated in May 2026 it is also trialling Ericsson’s AI-native scheduler with link adaptation on live 5G-Advanced traffic, but Soni noted AT&T has been trialling it since December 2025.

Soni told MWL AT&T is just past the halfway point for switching out Nokia’s RAN gear for Ericsson’s as part of a five year plan to transition 70% of its network traffic to run on open platforms by end of year.

The post AT&T scales cloud RAN as open RAN swap passes 50% appeared first on Mobile World Live.

The Linux Foundation launched the open-source Agentic AI Foundation (AAIF) six months ago to ensure protocols are standardised across enterprises, hyperscalers, and telecommunications companies.

The December launch spanned 49 companies, including co-founders Anthropic, Block, and OpenAI, with additional support from Google, Microsoft, AWS, Cloudflare and Bloomberg.

As of 3 June, it now has more than 200 member companies and is growing at twice the pace of the Linux Foundations CNCF Kubernetes project, AAIF executive director Mazin Gilbert told Mobile World Live.

“I think what is new that I have not seen before is that in every foundation I’ve been part of before, we used to go and ask, convince companies to join,” Gilbert said. “The phenomenon we’re seeing here is the opposite.

“Companies are coming on their own accord,” he added. “No outreach. They all want to have a say in this ecosystem.”

Being open source, all the members of AAIF benefit from the development of the tools and interfaces.

As the industry moves beyond chatbots and LLMs, Gilbert explained stakeholders need to agree on standard protocols and interfaces to deliver on agentic AI’s value.

Gilbert, whose career includes 22 years at AT&T and over four at Google, said there are roughly two million open-source models currently available, with open variants now only four months behind closed ones.

The infrastructure to run them is maturing fast but connecting those models to real applications, such as to self-healing networks, customer care systems and robotic platforms, requires standardised plumbing which nobody has built to date in a neutral, durable way.

“This foundation was created to ensure that the tools, the plumbing, the protocols are standardised,” Gilbert explained.

“I don’t need to know which language model out of the two million I should use,” he added. “I should be able to plug and play into any of them without any additional new code.”

He stated the real breakthrough will come from creating standard interfaces which enable AI agents to connect consistently with networks, databases, applications, microservices, other agents, and development tools.

When the plumbing is standardised, it will unlock a new ecosystem of interoperable agent-based systems, much like common internet protocols enabled the modern internet.

“We call it the internet of agents because now, suddenly, just like the internet, you can start to communicate, you can work with each other,” he said.

agentgateway
With the focus on plumbing, the AAIF welcomed agentgateway as its fourth hosted open-source project.

It is built on the A2A and MCP protocols and functions as what Gilbert calls a traffic light: routing agent-to-LLMa and agent-to-microservice communications, enforcing authentication, tracking token usage, and monitoring inter-agent channels.

For telcos, it offers a concrete path toward agentic network management without rebuilding their existing microservice estates from scratch.

“As far as the agents and the LLMs are concerned,” Gilbert said, “they look at the gateway as the endpoint.”

AI market
Setting aside the user application layer, Gilbert frames agentic AI as a four-layer stack: starting with AI-optimised hardware such as GPUs, TPUs and NPUs: moving up to AI-native infrastructure with Kubernetes, containers, service mech and data bases: then to foundation models, inference and model serving: and finally to the fourth agent applications layer where planning, reasoning and self-healing functions are expected to develop quickly over the coming months.

“We need to start populating what layer four is,” he said. “I expect the key work that we are doing now and over the next six months is to really make sure that we are looking at the agentic AI stack end-to-end.”

The roadmap for the next six months includes scaling 10 to 15 projects, a doubling of the working groups from the current eight, and the first production-scale deployments from member companies.

“We’re already starting to see some companies talking about thousands of agents already in production,” added Gilbert.

The post Analysis: How the Agentic AI Foundation is mobilising an agent army appeared first on Mobile World Live.