Why telecom operators are migrating to open AI models

Telecommunications operators are repositioning their artificial intelligence strategies over open models. The logic, according to NVIDIA, is direct: closed models tie the operator to a single supplier, a single cloud and a single set of rules on where the data can circulate. Open weight models allow you to scale, optimize and distribute workloads between public cloud, private infrastructure, and edge — and it is this flexibility that telecom companies require to sell AI to corporate and government customers.

Why do telecom operators prefer open AI models?

Two demands weigh: flexibility and security. Operators deal with network data, subscriber traffic and corporate contracts that, in many cases, cannot leave a jurisdiction or a data center of their own. Open models allow the operator to host the inference where it makes sense — at the core, in a public cloud contracted or at the edge, near the customer — without renegotiation of the model with each infrastructure change. NVIDIA describes this movement as a way to serve business and government customers with locally adapted AI services rather than generics.

Nemotron, NeMo and agents: the open AI model stack

NVIDIA points out three central pieces. Nemotron is the family of open models that serves as the basis for customization. NeMo is the training, adjustment and implementation environment that allows you to adapt these models to specific domains — telecom, service, networks. And AI agents are the layer that performs tasks: consulting systems, opening calls, automating operational flows. Together, they form what the company calls the base for operators who want to control their own AI stack.

The portal has already covered developments near this line: NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard shows the evolution of faster and more controllable agile models, and Mistral bet on sovereign AI with open weight models illustrates the same movement on the European side.

From data center to edge: where open AI models run

The architecture NVIDIA describes is hybrid by definition. Training charges stay in public cloud or in private clusters; sensitive inference runs on the operator's own infrastructure; and functions requiring minimal latency — real-time traffic analysis, network automation — go down to the edge. This division is not technical novelty, but the commercial argument is: the more open the model, the more freedom the operator has to move it between these environments without losing control over data and costs.

What changes to Brazil with open AI models?

For the Brazilian market, the discussion has two fronts. On the one hand, operators serving public agencies need to respond to LGPD and data sovereignty requirements — hosting an open model locally is simpler than depending on the policy of a foreign provider. On the other hand, the ecosystem of developers and integrators gains a layer of work: adapting open models to vertical domains. The theme of sovereign infrastructure has already been addressed here in Digital Soberania: How Countries Build IA Infrastructure(/noticias/soberania-digital-paises-infraestrutura-ia-prioridades-estrategicas).

What Open AI Model Ads Still Not Solve

NVIDIA describes the strategy, but does not publish adoption numbers, deployment costs or benchmarks comparing open models to closed alternatives in real telecom scenarios. There is also no detail of named customers or operators. Open models solve the problem of mooring to the supplier, but transfer to the operator the fine-tuning burden, quality assessment and responsibility on what the model responds — an engineering cost that does not appear on the slide.

Source: [NVIDIA Blog — Telecom Operators Build AI Strategy on Open Models.

Frequently Asked Questions about Open AI Models

What are open AI models in the context of telecom?

These are models with weights available for the operator to download, adjust and host wherever you want — in public cloud, private data center or edge — rather than rely on a closed API from third parties.

Why would an operator choose an open model instead of a closed one?

Mainly for flexibility and control: it decides where the data is, adapts the model to specific domains and distributes inference according to latency and cost.

What is the role of Nemotron and NeMo technologies in this strategy?

Nemotron provides base models, NeMo handles training, adjustment and deployment, and AI agents perform operational tasks on this basis.