MWC Barcelona 2026 was themed “The IQ Era.” After walking the show floor, sitting through sessions, and talking to engineers behind the booths, my takeaway is simple: the telecom industry has moved past AI experimentation.
Leading operators have moved beyond pilots and are now running AI agents in live production environments. The conversation has shifted from “if” this works to the gritty details of “how” to scale it. That “how” demands engineering depth that combines AI, networking, and systems thinking in ways that are genuinely hard to find.
Here’s what stood out, and how it connects to the engineering reality we see at CodiLime every day.
Agentic AI in the NOC is real, but the integration is where the most fail
The biggest theme at MWC 2026 was agentic AI, systems that don’t just assist operators but autonomously detect, diagnose, and remediate network issues.
- Microsoft reported
that one operator now runs 60% of NOC operations through AI agents, handling over 10,500 tasks/month. - Deutsche Telekom demonstrated MINDR , a multi-agent system that predicts risks and executes corrections autonomously.
- Rakuten achieved the world’s first TM Forum Level 4 autonomy validation for RAN energy efficiency, delivering ~20% energy conservation through autonomous operations on their cloud-native Open RAN architecture.
These are production deployments, not demos.
However, what the keynotes often don’t say: agents are only as good as the integration fabric underneath. Without a trusted source of truth, reliable topology and config state, safe automation APIs with guardrails and rollback, and telemetry pipelines that feed the right data to the right model, even the best model will fail.
We know this because we’ve built it. We develop AI-powered observability and troubleshooting systems that combine agentic workflows with real network data (topology, device state, syslogs, alerts) to accelerate root cause analysis and incident response. We’ve validated these approaches live with network engineers and in real operational environments. The industry is now converging on the patterns we’ve already been building.
Autonomous networks need orchestration, not just better algorithms
A major move at MWC was the announcement of cross-platform interoperability between Ericsson and Nokia of their rApp ecosystems. Telstra demonstrated self-healing networks recovering from outages in minutes. Google Cloud launched AI agents targeting Level 4-5 autonomy.
All of this requires network orchestration that actually works across vendors and scales.
We’ve been delivering exactly this. For one of the world’s largest technology companies, we built a network orchestration framework managing tens of thousands of network devices across data center fabrics and GPU clusters. For a major e-commerce platform, we built an intent-based orchestrator for thousands of devices across multiple SONiC-based fabrics, cutting provisioning from weeks to days.
The Ericsson-Nokia collaboration is a positive signal. But real networks run SONiC, Juniper, Cisco, Arista, and everything in between. Multi-vendor orchestration isn’t a feature you announce, it’s an engineering discipline you earn through delivery.
Edge AI and acceleration are now the default architecture
NVIDIA’s AI-RAN innovations tripled compared to last year. GIGABYTE showcased BlueField-3 DPU infrastructure . Mobilint demonstrated edge AI processors supporting 490+ deep learning models.
The message is loud and clear – AI inference is moving into the data path, and CPU-only processing can’t keep up.
This is our core territory. At CodiLime, we’ve deployed AI models on GPUs, DPUs, SmartNICs, and FPGAs for security and networking workloads:
- An AI-powered WAF on GPU + DPU with zero host CPU involvement, threats blocked before reaching the kernel, built for a network security vendor.
- ML-based SYN flood detection on FPGA SmartNICs at line rate, delivered for a leading semiconductor company.
- Embedded high-performance packet processing achieving millions of packets per second for a satellite communications provider, using DPDK with custom poll mode drivers.
When the industry talks about heterogeneous processing pipelines (CPU, NIC, DPU, FPGA, GPU), we’ve been stitching those together with real drivers, orchestration, and CI/testing across multiple projects.
Network security is the elephant in the room
As Palo Alto Networks announced “Secure by Design” AI Factories with Nokia and partners, and Huawei launched quantum-secure WAN with the industry’s first built-in Quantum Key Distribution (QKD) board in routers, the stakes have never been higher. A recent Cohesity survey revealed that 76% of organizations have experienced a material cyberattack, and telecom is no exception.
As networks become more autonomous, the attack surface grows. Security engineering that combines AI, networking, and software craftsmanship isn’t optional, it’s foundational.
We’ve been tackling this at scale – from security analytics platforms processing billions of netflows for a global telecom provider, through enterprise-scale XDR systems, to automation that has helped hundreds of organizations migrate and secure their network infrastructure. We also work on AI security itself, helping protect AI workloads from emerging threats.
Digital twins and test automation are now table stakes
If you are automating changes at scale, you need a safe place to fail. MWC confirmed that network engineering has officially become software engineering, with test frameworks, reproducible topologies, CI pipelines, and quality gates.
We’ve built digital twin environments and network labs for some of the largest infrastructure operators in the world. The pattern is always the same: if you can’t test it safely, you’ll never have the confidence to automate it.
My three takeaways from MWC 2026
1. The talent gap is the real bottleneck.
Every vendor is shipping AI agents, but few organizations have engineers who understand both networking protocols and production AI systems. NVIDIA’s open Nemotron 30B telco model is a starting point, not a solution. It needs domain-specific fine-tuning, integration with operational systems, and production hardening.
2. PoC-to-production is where most companies stall.
Deutsche Telekom’s message was clear: fewer demos, more scaled deployments. This transition, from notebook to production, from single-vendor to multi-vendor, from lab to operations, is exactly the engineering work we do every day.
3. Convergence is the story.
Telco, data center, and cloud infrastructure are merging around AI-driven requirements. Orchestration platforms, with a strong source of truth, lifecycle workflows, and multi-vendor integration, are the real battleground.
The IQ Era demands engineers who speak both AI and networks. That has been our thesis at CodiLime for a long time, and MWC 2026 just validated it at industry scale.
If you’re rethinking NOC operations, building AI-native platforms, or modernizing network infrastructure, let’s talk. We’ve likely solved a version of your problem already.
