- Red Hat sees open AI models helping telcos turn rising token costs into a new revenue opportunity
- AT&T, T-Mobile and Orange Business offer clues to how operators are rethinking AI costs, control and sovereignty
- Selling AI inference and agent services takes more than models
Operators are still figuring out how to make AI work for internal use cases at scale. Consumer-facing applications are even further and fewer in between. But open models could be a key ingredient when it comes to cooking up more of the latter even faster, Joe Fernandes, VP and GM of Red Hat AI, told Fierce.
“They are expressing an interest because it allows them to then actually provide these services on their own,” Fernandes said, when asked about telcos and open models. “So, if they want to provide inference and agents and then other network services, the way to do that on their own infrastructure is to be able to leverage their own models on that infrastructure.”
Rising token costs and demand for sovereign solutions are among the factors steering telcos down this road.
“People are more aware of token costs because they're consuming more and more and more of their budgets,” Fernandes told Fierce Network. As usage spreads across more employees and processes, he said, companies – including telcos – have started asking whether they should run models themselves. Or, as Fernandes put it, own instead of rent.
“At some point…you want to actually own that capacity and then become a provider,” Fernandes said.
Progress on the open front
Fernandes pitched open source as the “innovation engine that drives every new wave of technology,” though such a statement from a Red Hat executive is perhaps to be expected. After all, Red Hat’s entire business model is based on open-source technology.
Still, the open model approach is gaining traction among operators.
For instance, AT&T told Fierce in August its daily AI usage had climbed from around 8 billion tokens last year to upwards of 45 billion, while costs had stayed relatively flat thanks to a combination of open models, smart routing and fine-tuning. It expected to route 25% to 40% of its tokens through open models this year, with a broader goal of 65% to 70%.
Meanwhile, T-Mobile told Fierce that its model choices start with the workload’s performance, latency and reliability requirements. Its mix includes premium models for demanding customer experiences, smaller or open models for repeatable or data-heavy internal work, and traditional machine learning or automation where generative AI is not the right tool.
In France, Orange Business has opted to move sensitive AI workloads to open-weight models running on its infrastructure.
And telcos are far from the only ones warming up to open models. According to Vercel’s October 2026 AI Gateway index, open weight models now process more than 70% of AI tokens and account for 23% of AI spending.
Ultimately, Fernandes expects telco deployments to span multiple models, with different capabilities and sizes serving different tasks.
The catch: Operators have to run it
Leveraging open models, Fernandes sees an opportunity for operators to both gain more control over their own AI deployments and build the capability to offer products like model-as-a-service and agent management.
But running AI internally and selling it externally are distinct business cases. And taking more control also means taking on more operational responsibility. Operators need the skills to deploy and maintain the platform, as well as security controls suited to agents that act like applications, Fernandes said.
Model guardrails can filter incoming prompts and outgoing responses, but they do not cover everything an autonomous agent does. “You have to secure the agent itself, right?” Fernandes said. “And that includes not just calling the model, but also calling tools, calling potentially other agents.”
That’s where sandboxing, like Nvidia’s OpenShell, comes in. Sandboxing offers an opportunity to run agent processes, limit what they’re authorized to do and assign agent identities.
In terms of long-term management of new models and agentic systems, Fernandes said it’ll be up to telcos to adopt a mix of upskilling and partnerships.