- AWS telco chief Jan Hofmeyr thinks AI agents could be the secret to making telco APIs actually work
- He also talked up work on new Network Language Models for operators
- AWS is focused on software for the RAN currently but didn't rule out a silicon push, Hofmeyr added
AWS RE:INVENT, LAS VEGAS – Before they ever really got going, telco APIs veered off the road into a ditch and have seemingly been stuck there for years. But the advent of agentic AI could very well change that, for one big reason, AWS’ Head of Telecom Jan Hofmeyr told Fierce.
“The problem they have today is that everybody has to agree on the same standard. It’s not just the API, it’s how the API behaves. If you have an API for a specific function, it has to behave exactly the same across any network,” he said. “I think agents can reason about that, but APIs cannot.”
The whole idea behind network APIs is for operators to expose their network capabilities and allow developers to create new applications using those raw materials. However, there have been two key problems: no two networks are created equal, making it hard for APIs to work across networks. And operators have had trouble tapping into a large-scale developer base.
There have been attempts to address the latter problem in fits and starts. But the APIs themselves have remained a hot mess for more than two years now. That makes Hofmeyr’s suggestion – or, honestly, anything that could move the ball forward – a very interesting proposition indeed.
His vision goes something like this: with the rise of agentic AI, the focus shifts away from the network APIs themselves and more toward making an agent available in the network that can act on behalf of the system. In other words, the API will sit behind an agent that will be the one responsible for exposing the system’s capabilities to developers.
Hofmeyr said APIs will clearly still have a role to play. They still need to be there because agents and MCP servers will still need to call the APIs. But the need for them to be entirely consistent across networks is reduced, he said.
Building network agents
The hurdle then becomes building competent network agents. On this front, Hofmeyr said AWS has started working with operator partners to build something he called Network Language Models.
To be clear, Hofmeyr said what he’s talking about is different from what SK Telecom has done in creating a Korean LLM. It even sounds different from what SKT and Deutsche Telekom set out to do in creating an LLM designed specifically for telecom customer service.
Hofmeyr described Network Language Models as LLMs that are distilled into small language models that have expert knowledge of an operator’s unique network design, elements and operations.
Though it sounds like these might take a while to get right, Hofmeyr said the process will be exponentially faster using tools like AWS’ newly unveiled Bedrock Reinforcement Fine Tuning capability.
“I think it will happen very quickly,” he said. “If I have to predict, it will be probably within weeks or months that we will have our first models and then we’ll just refine it.
AWS in the RAN?
There was one other thing we couldn’t resist asking Hofmeyr about in light of Nokia and Nvidia’s recent partnership: AI-RAN. More specifically, we asked whether or not AWS might be interested in deploying its silicon into operator networks.
Hofmeyr said AWS’ current plan is more software-focused but didn’t rule out the possibility.
“We are very much focusing on EKS [Amazon’s Elastic Kubernetes Service] as the CaaS [Container as a Service] layer on which you run your AI agents because it makes it agnostic to the hardware,” he said.
“It’s too early now for us to decide what’s the right hardware that will enable this in the future now. My intent is that we enable it on today’s hardware,” he continued. “We will over time see if there’s an opportunity for us to bring silicon to the RAN.”