NVIDIA pushes AI agents into telecom networks at DTW Ignite 2026

NVIDIA used TM Forum's DTW Ignite 2026 in Copenhagen to lay out a full agent stack for telecom operators, with SoftBank, AdaptKey, Amdocs, NTT DATA, ServiceNow, TCS, Forsk, VIAVI, and KDDI on stage. The pitch: agentic networks need more than a chat model. They need synthetic data, domain-tuned models, a secure runtime, and accelerated simulation, so an operator can hand a long-running agent a real job, per the NVIDIA blog post.

What NVIDIA and telecom partners are showing at DTW Ignite

The NVIDIA blog post published on June 22, 2026 frames the move in one line. Automation is no longer the finish line, it is the launchpad to autonomy. That is more than a slogan. It is a description of the gap between the task-based generative AI most carriers have already shipped, and the always-on, multi-step agents that the next generation of network operations will need.

Operators are already deep into the first wave. Generative AI has cut time on network management, customer care, and back-office work, but most of that impact is a faster version of the same hand-correlated workflow. People still gather the alerts, write the runbook, and tell the next tool what to do. The DTW Ignite work is about closing that loop, so an agent can watch the network, decide what is wrong, and either fix it or escalate it with the right context attached.

The agent stack is a different engineering problem

Three pieces of the stack are the reason this is not just another AI feature release. The first is data. 54% of operators cite data-related issues as their biggest barrier to AI rollouts, because the most valuable network and customer data is also the most sensitive. The second is the runtime. Long-running agents that respect service-level agreements, change-management policies, and regulatory constraints are a different engineering problem from a customer-service prompt. The third is validation. Before an agent touches a live network, operators want to know that its recommendations are safe, and that means accelerated simulation running in the same loop.

SoftBank Corp. is the visible example for the data layer. The operator is using NVIDIA NeMo Safe Synthesizer and NeMo Anonymizer to generate privacy-preserving datasets that match the structure and distribution of real network performance and configuration data. Those synthetic datasets feed a large telecom model that SoftBank is fine-tuning for specialized network agents, and the same data pool is available to internal teams and external developers without exposing raw customer records.

The runtime layer is where the partner list gets long. NVIDIA NemoClaw blueprints and the NVIDIA OpenShell secure runtime give the agents policy-based guardrails and sandboxed access to telecom systems, so the operator can expand what an agent is allowed to do without losing auditability. AdaptKey is piloting security-hardened, long-running agents for self-healing 5G network operations on top of NemoClaw and OpenShell, with KeySmith orchestrating diagnosis and the auditable fixes that follow across core, radio access network, and billing systems. Amdocs is showing the same runtime applied to proactive customer-care agents, including a roaming scenario where an agent sees a customer's roaming package is about to run out, reaches out with the approved options, and executes the action inside defined business policies. NTT DATA is pairing NVIDIA Nemotron open models with NemoClaw to build long-running agents that track long-term performance trends and escalate the cases that need deeper telemetry analysis. ServiceNow is bringing Project Arc into the telecom network operations center, where it pulls context from emails, logs, and diagnostics across systems that do not normally talk to each other and runs the full incident response from first alert to assigned work order. TCS is layering a multi-fidelity AI sensor architecture over NemoClaw and Nemotron, with NVIDIA NV-Tesseract models that scan broadly for issues and trigger deeper diagnosis only when something looks off.

The day an agent can fix the network itself

The simulation story is the part that lets the rest of the stack ship. If an agent can validate its plan against a digital twin of the live network before it acts, the operator can let it run with weaker human-in-the-loop checks. Forsk has put an AI-based radio propagation model inside 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. VIAVI is doing the same thing with its TeraVM AI RAN Scenario Generator, with early results showing order-of-magnitude improvements in simulation throughput, and a new IP Network Configuration Blueprint that extends the same validation to the IP and transport layers. KDDI and KDDI Research are pushing this into 6G research with Keysight and Samsung Research America, building a high-fidelity RAN digital twin on NVIDIA Aerial Omniverse Digital Twin so multiple autonomous agents can run what-if scenarios against future radio conditions and traffic shifts at the same time.

The pitch to operators is that the four pieces of the stack, synthetic data, telecom-domain reasoning models, a secure runtime, and accelerated simulation, are what turn a copilot into an autonomous agent. The DTW Ignite 2026 showcase is the first time NVIDIA has pulled all four into a single partner lineup, and the operator names on the list are not the small carriers. SoftBank, NTT DATA, KDDI, TCS, and ServiceNow are the kind of partners that signal to a chief network officer that the agent stack is ready for a real procurement conversation, not just a proof of concept.

There is still a gap between a demo floor in Copenhagen and a carrier turning on autonomous operations in production. The 54% data barrier is real, the policy engines that NemoClaw and OpenShell depend on are still being tuned to local change-management rules, and the digital-twin loop has to be calibrated against a real network before an agent can be trusted to act on it. What has changed is that the building blocks are now in the same place, from the same supplier, with the same partner ecosystem behind them. For an industry that has spent the last two years stitching agent features together from five different vendors, that is the part of DTW Ignite 2026 that actually matters.

For a broader look at how AI infrastructure is being built out across the agent stack, see the AI Infrastructure in 2026 reference page.