Why Telecom Operators Are Choosing Open AI Models

👁 113 reads

Closed AI models from major providers are expensive, difficult to control, and - when you deploy them in critical telecom infrastructure - raise legitimate regulatory concerns. The telecom industry has reached this conclusion collectively: 89% of respondents in NVIDIA’s State of AI in Telecommunications report say open models and software are important to their company’s AI strategy. This isn’t fashionable enthusiasm - it’s an architectural decision.

Five reasons telecom chooses open source

In its latest analysis for the telecommunications sector, NVIDIA lists five strategic reasons why operators are turning to open weights. First is access to frontier models at lower cost - closed systems are reserved for narrow use cases where they truly pay off. Independent benchmarks, such as Artificial Analysis Intelligence Index v4.3.2, show that leading open models are catching up to commercial rivals in reasoning, coding, scientific tasks, and agentic workflows.

The second reason is the ability to customize for telecom specifics: open weights and training recipes allow fine-tuning models on proprietary network, customer, and operational data. Third - transparency and control over model behavior, which directly translates to regulatory compliance. Fourth - deployment flexibility: public cloud, private infrastructure, edge - without vendor lock-in. And fifth, the most commercially interesting: operators can host and fine-tune models themselves, then sell them to corporate and government clients as locally adapted AI services.

This isn’t a wishlist. It’s a description of a real business model that’s starting to work in practice.

SoftBank builds its own telecom model

The best proof is SoftBank Corp. The operator uses open foundations - including NVIDIA Nemotron models - to build its own SoftBank Large Telecom Model, designed for network operations, design, and infrastructure management.

“Open models allow SoftBank Corp. to build on the rapid progress of global foundation models while applying the network knowledge and operational expertise we’ve accumulated over many years”

- said Rajeev Koodli, principal fellow at SoftBank Corp. and senior vice president at SB Telecom America. This statement is worth reading carefully: SoftBank isn’t replacing global models with its own - it’s augmenting them with domain knowledge that no general model has or will have.

This is precisely the architectural value of open source in telecom. Network operations aren’t a customer service chatbot - precision matters in router configuration, protocol knowledge, understanding how a specific operator’s specific network is built. No model trained on internet data knows what SoftBank’s network topology looks like in Tokyo. But SoftBank knows - and can transfer that knowledge to an open model through fine-tuning.

30 billion parameters specifically for telecom

NVIDIA simultaneously announced a concrete product: Nemotron 3 Large Telco Model with 30 billion parameters, fine-tuned by AdaptKey on open telecom datasets. The model understands industry terminology and can reason through typical workflows - network configuration, customer incident handling, ticket triage.

Additionally, NVIDIA published the complete fine-tuning recipe - an end-to-end pipeline based on NVIDIA NeMo libraries - that allows operators to adapt Nemotron 3 LTM or other open models to their own networks, customers, and procedures. This is significant: it’s not about a ready-made product you buy and deploy, but a methodology you can replicate on your own data.

When openness becomes competitive advantage

For years, the industry intuition was that “the best model is the one with the most parameters from the biggest provider.” This intuition is slowly changing - not because large models are getting worse, but because open models are becoming good enough while offering something closed systems cannot: ownership rights to data and model artifacts, the ability for regulators to audit, control over where the model runs.

For an operator managing critical infrastructure, these aren’t nice extras. They’re necessary conditions.

It’s also interesting that NVIDIA - a company that profits from hardware for training and running models - actively promotes open ecosystems. The logic is simple: the more operators build their own models on open foundations, the more GPUs they need for fine-tuning and inference. For NVIDIA, open source isn’t ideology, but market strategy - and this particular alignment of interests works in the entire industry’s favor.

Telecom operators rarely pass for technology pioneers. This time they seem to understand something many other sectors haven’t yet absorbed: in AI, control over the model means control over the business.

// read moreNVIDIA Opens Vera Rubin Architecture to Scale Gigawatt AI FactoriesNewsWhy AI Open Source Is Strategy, Not Altruism: AstaBriefMagazineThere’s No Such Thing as Free AI: Who Really PaysMagazine

Kinetyra