- T-Mobile is adding open AI models to its enterprise AI stack,
- The operator says model choice starts with use case, not just cost
- Model routing and AI failover are becoming key new telco disciplines
Open models are making their way into T-Mobile’s enterprise AI stack, but the move isn’t just about cost. It’s about matching the right workload to the right model as part of what the operator has dubbed its “fit-for-purpose” AI strategy.
“Cost and control matter, but our choice starts with the use case: who we are serving, what experience we are trying to create, and what level of performance, latency and reliability the workload require,” a T-Mobile spokesperson told Fierce.
T-Mobile is using a mix of frontier and open options from various providers and is taking a pragmatic approach to model routing. For instance, it taps premium models for latency- or performance-sensitive customer experiences, smaller or open models for repeatable or data-heavy internal work, and traditional ML or automation where GenAI is not the right long-term tool.
“Open models are getting stronger and could make sense for more use cases over time, especially routine or repeatable work. But for customer-facing use cases like real-time voice, proprietary models are still ahead today,” the spokesperson explained.
What do telcos think about open AI models?
Nvidia’s recent telco AI survey found 89% of operators said open-source models and software are important to their AI strategy. As more models work their way into the equation, AI model routing is becoming a new operational discipline telcos have to master. And each has their own calculus for deciding what task is sent where.
For instance, AT&T recently told Fierce it developed its own Smart Router to decide whether to send a given task to a premium model or a cheaper, open option. Among thousands of decision benchmarks, AT&T’s router factors in the cost to run each model in different cloud environments and whether or not a cache is involved in the request.
For T-Mobile, though, “routing is not just about cost” but also factors like task complexity, reliability, latency, performance and the level of control needed. The rep pointed to the operator’s efforts to build a self-optimizing, self-healing network as an example of how it chooses different models for different work.
“In situations like a natural disaster, AI can help analyze network data and identify adjustments that help reduce coverage gaps and keep customers connected. For those kinds of operational use cases, speed, reliability and access to real-time data matter as much as cost,” the rep said.
How can telcos manage risk?
T-Mobile’s model routing calculus also has to account for what happens when things break. That’s because while not common, cloud outages do happen. And operators can’t afford to have their cloud-hosted models – and the applications those run – go down with the ship.
For its part, T-Mobile’s rep said it manages AI failover much like any other resource, using redundancy and fallback strategies to reduce risk. “In some cases, we can fail over to cloud models at the application layer, but as our on-prem capability matures, we expect that will become less important,” the rep said.
That resilience becomes more important as AI moves into more operational environments. The goal, T-Mobile said, is to use AI to support faster, smarter decisions without turning any one model — or the infrastructure behind it — into a single point of failure.
Read more about how telcos are using open AI models here:
Open models are driving AT&T’s AI ‘tokenomics’ strategy
GSMA warns telcos against outsourcing AI future to hyperscalers
Open-weight AI cuts costs and lock-in — with a catch
GSMA seeks to tailor AI models for telco requirements
OpenAI's new open weight models could change game for telcos