Breaking

NVIDIA Says AT&T, SoftBank and Indosat Are Adopting Open Models for Telco AI Agents

October 6, 2026 at 09:13 EDT

AI Agents
Foundation Models
Infra & Chips

NVIDIA is positioning a full-stack approach for turning open models into controlled telecom AI agents, saying AT&T, SoftBank and Indosat are adopting open models for use cases including autonomous network operations. The push puts domain-tuned models, synthetic data, agent harnesses, sandboxes and network simulations at the center of carriers' move from automation to autonomy.

Oct. 2026 · NVIDIA and telcos

Open Models Become Telco Agents

AT&T, SoftBank and Indosat are moving domain-tuned open models toward governed network agents, with NVIDIA supplying model, runtime and infrastructure pieces for controlled autonomy.

45B

AT&T tokens handled per day

400B

tokens used to post-train OTel 2.0

1 GW

Zankore DSX AI Factory target

operators named in the adoption push

Fine-tuning Lifted NOC Accuracy

A telco reasoning pipeline improved incident summary accuracy from roughly 20% to 60%.

20%

Baseline Qwen3-32B

60%

Fine-tuned telco model

How the Agent Stack Works

Open model

Nemotron LTM, OTel, LTM or Sahabat-AI

→

Telco tuning

Logs, standards, procedures and synthetic data

→

Agent harness

Plans, tool calls and network context

→

Sandboxed action

OpenShell, NemoClaw and operator policy

Why operators like it

Open, domain-tuned models can run closer to carrier data, lower token costs and keep operators in charge of policy.

What still blocks it

Sensitive data, model accuracy, scarce AI expertise and safe production change control remain hard problems.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.