As AI becomes increasingly embedded in everyday telecom operations, inference has become a cost that grows with every token. Telecom-trained models offer operators a way to bring that cost down without giving up accuracy.
For the past few years, the telecoms industry’s AI question has been about capability: which model is the smartest, and how quickly can it be put to work? But as AI moves from pilots into daily operations, the bill is starting to matter as much as the benchmark.
According to TM Forum, AT&T now runs more than 40 billion tokens on an average day, up from 5 billion in September 2025. At that volume, inference becomes an operating cost that needs managing in the same way as network capacity or cloud spend.
It also changes what operators should ask of their models; it’s no longer just about which model is smartest, but which model is good enough for the task at the lowest cost-per-answer.
General-purpose models pay a context tax
Most operators today use general-purpose models from the large AI labs, supported by Retrieval-Augmented Generation (RAG) to search internal documents and industry standards whenever the model needs more context.
This works, but it comes at a price.
Every time a general model is handed the 3GPP specification or TM Forum standard it should already understand, the operator pays twice: in the tokens needed to supply that context, and in the latency of retrieving it.
Call it the context tax. On occasional queries, it barely registers, but across billions of daily requests, it becomes one of the largest avoidable costs in an operator’s AI estate.
Gartner notes that the accuracy of general-purpose models declines on tasks that need specific business domain context, and telecoms is about as specialised as domains get. The GSMA reports that just 16% of telecoms generative AI deployments have been applied to network operations, a gap it links to the difficulty general models have with network data and standards documentation.
Domain models are now winning on telecom tasks
That gap is starting to close. AT&T’s OTel 2.0, developed through the GSMA’s Open Telco AI initiative, was post-trained on 400 billion telecom-specific tokens selected from more than a trillion processed. Its base dataset draws on seven standards bodies, including 3GPP, ETSI, ITU, O-RAN and TM Forum.
The results are hard to ignore. In the GSMA’s July 2026 assessment of the Open Telco AI benchmarks, the top three performers were all domain-adapted models rather than general-purpose ones. The wider Open Telco AI initiative also extends beyond AT&T: SK Telecom, SoftBank, Orange, KDDI and Swisscom are contributing to the same initiative while developing telco-specific models of their own.
It fits a wider shift. Gartner predicts that by 2027, organisations will use small, task-specific AI models at least three times more than general-purpose large language models.
Shared industry knowledge, operator-specific advantage
A fair question is why a model trained by AT&T should matter to anyone else. The answer is that models like OTel are trained mainly on shared industry knowledge, the standards, protocols and architectural principles every operator already works with, rather than on any one operator’s data.
That makes them a starting point, not a finished product. A telecom model won’t know a particular operator’s product catalogue or charging rules out of the box. That knowledge still has to come from the operator’s own data, processes and governance. But fine-tuning a model that already understands the domain is a far smaller job than teaching a general model telecoms on every request.
Open models also give operators more choice over where AI runs, whether in public cloud, private cloud or on premises. That matters for sovereignty, for regulatory control and for avoiding dependence on any single AI provider.
Orchestration, not replacement
None of this means abandoning frontier models. They remain the better choice for tasks that need broad knowledge or deep reasoning. The point is that most operational tasks don’t.
AT&T’s Chief Data and AI Officer, Andy Markus, puts it plainly: “only a small percentage of the tasks we run require that level of sophistication.” The operator has built a gateway that routes each task to the most cost-effective model capable of handling it, and says the approach is reducing AI costs by as much as 90%.
The lesson for other operators is that model choice is becoming an architectural decision, not a procurement one. The aim is not to find one model that does everything, but to match each task to the right model on accuracy, cost, security and control.
Where telecom-native models will land first
The strongest early fit is the high-volume, repeatable work of BSS/OSS, where domain knowledge matters and the same types of task recur thousands of times a day:
- interpreting charging rules
- investigating mediation issues
- explaining billing anomalies
- supporting network configuration
- working with TM Forum Open APIs and telecom standards
Here, AI can act less like a generic assistant and more like a telecom specialist. But specialism does not remove the need for control. AI models are probabilistic; billing and provisioning cannot be. Any action an agent proposes, such as moving a customer to a new tariff, should run through governed interfaces such as TM Forum Open APIs, so existing systems can check eligibility, pricing, authorisation and compliance before anything changes.
Three reasons for caution
Telecom-native models are not a guaranteed win.
- The industry is not uniform. Standards are shared, but implementations, regulation and languages vary by market. A model trained mainly on English-language standards may need significant local adaptation.
- The knowledge base moves quickly. 3GPP releases, TM Forum APIs and regulation change constantly, so keeping a telecom model current is an ongoing engineering commitment.
- Frontier models are improving. Each new generation may narrow the domain knowledge gap, reducing the advantage of specialised models over time.
Managing models like capacity
Operators already know how to run a portfolio of resources against changing demand. They do it every day with spectrum, network capacity and cloud infrastructure. AI models are joining that list.
The AI-native operator will not run on a single all-purpose model. It will run on a governed mix of specialised and frontier models, each matched to the work it does best. The winners will not be the operators with the most powerful model. They will be the ones paying the least for the right answer.