Why Telecom Operators Are Building Their AI Strategy on Open Models: NVIDIA Releases a 30B Nemotron 3 Telco Model and Fine-Tuning Recipe

What was announced

On October 6, 2026, the NVIDIA Blog published an article by Kanika Atri titled "Why Telecom Operators Are Building Their AI Strategy on Open Models." The central claim is simple. Telecom operators increasingly base their AI strategy on open models, and the reasons go beyond cost. Open models let operators trust, control and customize AI across their most critical workloads, from autonomous networks to customer care. NVIDIA's latest State of AI in Telecommunications report supports the claim: 89% of respondents said open source models and software are important to their company's AI strategy.

One caution applies to everything below. This is a vendor blog. The survey figure, the partner quotes and the product announcements all come from NVIDIA's own account. We treat the piece as a useful industry signal, not as neutral third-party research, and we separate what the source states from what we infer.

The five strategic benefits of open models

The article lists five benefits for operators. First, open models widen access to frontier-level intelligence at lower cost, so operators can reserve closed models for the workloads where they add the most value. The article cites independent benchmarks such as the Artificial Analysis Intelligence Index v4.3.2, which show leading open models becoming more competitive on reasoning, coding, scientific and agentic workloads. Second, open weights and training recipes support telco-specific customization: operators can fine-tune on their own network, customer and industry data.

Third, open models enable trustworthy AI. Operators get more visibility into, and control over, model artifacts and behavior, so they can evaluate, adapt and govern models in line with regulation and business policy. Fourth, deployment is flexible and secure. Teams can size and optimize open models to run on public clouds, private infrastructure and edge environments. Fifth, operators can host and fine-tune open models to deliver locally adapted AI services to enterprise and government customers.

In our analysis, the first and third points are the most structural. The first describes workload-level model routing: neither all-open nor all-closed, but tiered by the value of each workload. The third is about governance. Telecom is a heavily regulated industry. Data sovereignty, auditability and explainability often matter more than a small gain in raw accuracy, and open weights have a natural advantage on those dimensions.

Technical and product details: Nemotron and Nemotron 3 LTM

The NVIDIA Nemotron family of open models offers frontier-level reasoning tuned for agentic workflows, plus speech capabilities for voice applications. It ships with open weights, open training data and open recipes. That goes further than releasing weights alone, because an operator can reproduce and modify the training process itself.

The article announces the 30-billion-parameter Nemotron 3 Large Telco Model (LTM). AdaptKey fine-tuned it on open source telecom datasets to improve accuracy on telecom-specific tasks. It gives operators an open baseline that understands telecom terminology and can reason through operational workflows such as network configuration and customer incident triage.

To help operators customize Nemotron 3 LTM and other open models with their own operational data, NVIDIA released the full recipe for the end-to-end fine-tuning pipeline. It uses the NVIDIA NeMo open libraries and adapts open models to operator-specific networks, customers and procedures. One gap is worth noting: the source gives no benchmark scores for Nemotron 3 LTM and no size for the accuracy gain. It says only that accuracy improves. Buyers should ask for numbers.

From models to production workflows

The article stresses that open models are a critical building block, but that models alone do not bring autonomous telecom operations safely into production. Operators need pipelines that prepare and protect data for fine-tuning, by anonymizing sensitive records and generating privacy-preserving synthetic datasets. They also need a platform that turns open models into governed, autonomous agentic workflows.

NVIDIA's answer is an end-to-end platform powered by NVIDIA AI Enterprise software and the NVIDIA Agent Toolkit. It spans data pipelines, open models, agent orchestration, secure runtimes and simulation, with partners building at every layer. Read plainly, the commercial logic is that the open model is the door and the software stack and accelerated computing are the room behind it.

Operator perspectives

SoftBank Corp. is using open foundations extensively, including NVIDIA Nemotron models and others, to develop its SoftBank Large Telecom Model. Rajeev Koodli, principal fellow of SoftBank Corp. and senior vice president of SB Telecom America, said open models let the company build on the rapid progress of global foundation models while applying years of accumulated network knowledge and operational expertise. The target use cases are network operations, design and overall management.

Andy Markus, chief data and AI officer of AT&T, said the future of AI is not about choosing a single model. It is about matching every workload to the right combination of performance, cost and control. He called open models essential to that approach and said AT&T wants to bring the model-choice strategy into production with the scale, reliability and governance its business requires.

Chirag Sukhadia, chief data and AI officer of Indosat Ooredoo Hutchison, one of Indonesia's largest operators, said the value of open models for a country like Indonesia goes beyond access to powerful AI. It is about adapting that intelligence to local language, culture, data and real-world needs. Indosat puts this into practice through its Sahabat-AI family of open source models.

What this means for developers and enterprises

For operator AI teams, the starting line moves forward. With an open telecom baseline and an official fine-tuning recipe, a team does not need to pretrain from scratch. It can spend its effort on data governance and evaluation instead. For enterprise and government customers, an operator that hosts and fine-tunes open models can become a provider of locally sovereign AI, meeting needs for in-country data, local language and sector compliance.

For the wider developer ecosystem, the article points to an emerging pattern for vertical industries: an open foundation model, industry fine-tuning, and a governance platform around both. Open weights, open data and open recipes together lower the barrier to reproducing an industry model.

Challenges and outlook

Four challenges stand out. The first is evidence. The 89% figure comes from NVIDIA's own report, and the article does not disclose the sample size or method. The second is that open does not mean safe by default. More visibility brings more responsibility for evaluation, red-teaming and continuous governance. The third is data. Fine-tuning needs high-quality network and customer data, which is highly sensitive, so the quality of anonymization and synthetic data will shape results. The fourth is autonomy itself: agentic workflows that touch live networks need rollback paths and clear failure metrics.

Looking ahead, as national AI strategies push operators to build local AI infrastructure, open models are likely to become one default option for telecom AI, with closed models kept for a small set of high-value uses. NVIDIA GTC Berlin runs October 20 to 22 and may bring more deployment disclosures. Two things are worth watching: public evaluations of Nemotron 3 LTM, and real-world failure rates for autonomous network agents.