Nvidia used TM Forum's DTW Ignite 2026 in Copenhagen this week to push telecom operators past task-level automation and toward fully autonomous networks, unveiling a stack of synthetic-data tools, telecom-domain models, secure agent runtimes and GPU-accelerated simulation. The headline pieces are NVIDIA NemoClaw, a blueprint for long-running agents, and NVIDIA OpenShell, a sandboxed runtime designed to keep those agents inside policy. SoftBank Corp, Amdocs, NTT DATA, ServiceNow, Tata Consultancy Services and AdaptKey are all building on the stack. The pitch is direct: 24/7 agents that can stick with a complex job from alert to resolution, not scripts that need a human to chain them together.
The data problem is the gating issue. Nvidia cited a figure that 54% of operators flag data-related issues as their biggest barrier to building telecom reasoning models, because the most useful network and customer data is also the most sensitive. The company's answer is synthetic data generated by NVIDIA NeMo Safe Synthesizer and NVIDIA NeMo Anonymizer, which produce datasets that mirror the structure and distribution of real network performance and configuration data without exposing raw records. SoftBank Corp is using both to fine-tune its large telecom model and train specialized network agents.
NemoClaw and OpenShell are the runtime layer. Together they give agents policy-based guardrails, sandboxed access to telecom systems, and an auditable trail of every action — the prerequisites for letting software touch a live carrier network governed by service-level agreements, change-management rules and regulators. Nvidia is framing this as the difference between an agent that runs a single diagnostic command and one that owns an incident end-to-end.
Key facts
- 01Nvidia launched NemoClaw blueprints and the OpenShell secure runtime to give telecom AI agents policy-based guardrails and sandboxed system access.
- 02SoftBank Corp is using NVIDIA NeMo Safe Synthesizer and NeMo Anonymizer to generate privacy-preserving synthetic datasets for fine-tuning its large telecom model.
- 03Forsk's Naos RAN platform hit ray-tracing accuracy up to 200x faster than CPU baselines on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.
- 0454% of telecom operators cite data-related issues as the top barrier to building telecom-domain reasoning models.
- 05Amdocs, NTT DATA, ServiceNow, TCS and AdaptKey are all building long-running agents on the Nvidia stack for care, network ops and self-healing 5G.
AdaptKey is piloting security-hardened agents for self-healing 5G operations, with NemoClaw and OpenShell powering detection agents that submit scoped remediation requests into AdaptKey's KeySmith platform. KeySmith then orchestrates diagnosis and applies fixes across core, RAN and billing systems. Amdocs is using the same runtime for proactive customer-care agents that, for example, spot customers whose roaming package is nearly used up, present approved options and execute the upsell within defined business rules. Amdocs is also pointing the stack at data-science agents that rank customers by readiness to migrate to modern billing platforms.
NTT DATA is building long-running agents on NVIDIA Nemotron open models and NemoClaw to track network-degradation trends and hand off cases to research agents for telemetry analysis. ServiceNow is bringing Project Arc to telecom, with autonomous network-operations-center agents that pull context from emails, logs and diagnostics, then run the full incident lifecycle from alert to work order. Arc actions are secured by OpenShell and governed by ServiceNow AI Control Tower. TCS is wiring NemoClaw together with Nemotron and NVIDIA NV-Tesseract to build a multi-fidelity "AI sensor" architecture that scans broadly and triggers deeper diagnosis only where it's warranted.
Simulation is the trust layer. Before an agent acts on a live network, operators want it to rehearse the change against a digital twin and show its work. Nvidia is pushing those workloads onto GPUs to make rehearsal fast enough to be useful in production. Forsk has integrated an AI-based radio propagation model into its Naos RAN planning platform, hitting ray-tracing-level accuracy up to 200x faster than CPU-only baselines on NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. The resulting RAN digital twin enables near-real-time use cases like network self-healing and automated antenna tilt.
VIAVI Solutions moved its TeraVM AI RAN Scenario Generator from CPUs to the same Blackwell Server Edition GPUs and reported order-of-magnitude improvements in simulation throughput, letting operators run high-fidelity scenarios at real deployment scale. VIAVI also released an IP Network Configuration Blueprint that extends validation into IP and transport, so routing and traffic-engineering changes can be tested before they hit production. KDDI and KDDI Research are going further out, partnering with Nvidia, Keysight and Samsung Research America to build a 6G-era RAN digital twin on NVIDIA Aerial Omniverse Digital Twin running in KDDI's AI data centers, where multiple autonomous agents will simulate area-optimization, traffic shifts and AI air-interface functions.
The skeptical read is that none of this is a deployed autonomous network yet — it's a stack of blueprints, pilots and demos at an industry conference. Telecom is one of the slowest enterprise verticals to adopt new runtime layers, and "agents touching billing and RAN under policy" is exactly the kind of change that regulators, internal risk teams and union-covered ops staff will move on carefully. The 54% data-barrier number is itself an admission that most operators are still upstream of the model-training step, never mind the autonomous-agent step.
For Nvidia, the strategic logic is straightforward. Telecom is one of the few verticals with the data volume, the simulation workload and the regulatory complexity to consume Blackwell GPUs at the scale Nvidia's roadmap requires, and it's a market where systems integrators like Amdocs, NTT DATA and TCS already own the customer relationship. By shipping the runtime (OpenShell), the orchestration blueprint (NemoClaw), the models (Nemotron) and the simulation environment (Aerial Omniverse Digital Twin) as a single stack, Nvidia is doing in telecom what it has already done in robotics and life sciences: define the reference architecture before competitors can. The carriers get a faster path to autonomous ops; Nvidia gets the silicon socket for the next decade of network buildouts.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
