By: Pavan Thalak

Telecom operators have stopped debating whether AI belongs in their business. In NVIDIA's fourth annual State of AI in Telecommunications survey, 90 percent of operators reported that AI is already helping them grow revenue and reduce costs, and 89 percent said they plan to raise their AI budgets this year, up from 65 percent the year before. More than a third expect those budgets to grow by more than 10 percent. The capital is committed, and the conviction behind it is hard to miss.

What's far less settled is whether that spending translates to the business results operators were promised. Plenty of pilots look impressive in a controlled demo, then stall on the way to production. The distance between an AI budget and an AI outcome has become the defining question of this moment, and it has little to do with how much an operator spends.

The industry settled the ‘why AI?’ question, but then got stuck on ‘how’ 

The appetite for AI is genuine and broad. NVIDIA's data shows that 60 percent of operators now use or are assessing generative AI, up from 49 percent in 2024, and 65 percent report that AI already drives automation across their networks. 

Network automation has overtaken customer experience as the top investment priority, with half of operators naming autonomous networks as their best-performing use case for return on investment. Looking further ahead, 77 percent expect AI-native networks to arrive before 6G, and a growing number of operators have started to describe themselves less as carriers of traffic and more as infrastructure companies for intelligence.

The momentum reaches the newest tools, too. Telecom posted the highest adoption of agentic AI of any industry in Nvidia's research, at 48 percent, and 89 percent of operators flagged open-source models as important to their strategy. Productivity gains are nearly universal, with 26 percent of operators reporting major improvements. The technology works, and operators know it works.

Readiness hasn't kept pace with that ambition 

TM Forum's research into generative AI maturity tells the other half of the story, and the numbers are sobering. Only 25 percent of operators feel equipped to use advanced techniques like fine-tuning and retrieval-augmented generation, and just 16 percent feel confident using those methods to manage cost and prove return. About a third have a senior executive dedicated to AI strategy. Only 14 percent run more than 10 generative AI use cases in production, and roughly a third report a mature pipeline of work behind the ones they've already shipped.

The most common reason initiatives stall is accuracy, which tends to break down once a proof of concept meets the messiness of live operations. TM Forum reports that operators keep trying to attach generative AI to systems that were never designed for it, rather than building intelligence as a native part of how the business runs. That one architectural choice explains far more of the variance in results than the size of any budget does.

Architecture decides whether AI compounds or stalls 

When AI sits on top of fragmented legacy systems, every use case turns into its own integration project. A customer-service bot might connect to one set of records while a fraud model connects to another, and neither shares context or learns from one another. What an operator ends up with is a portfolio of disconnected pilots that each deliver a thin slice of value while the cost of holding them together keeps climbing. The pilots that succeeded during the demo never compound into anything larger, because the architecture underneath them was never built to let them.

Embedding AI at the core of a unified stack helps solve this problem. The same intelligence layer that resolves a customer's billing question can recognize that the customer is a strong candidate for an upgrade without handing off to a separate system. Value compounds because customer experience, monetization and operations all draw on one model of the customer instead of three competing versions. A gain in one area feeds the next rather than stopping at a system boundary.

AI-native architecture creates value across three critical domains 

The impact of AI-native architecture extends well beyond network operations. Operators investing in the core architecture are creating advantages across three areas: consumer experience, developer efficiency and operational excellence.

For consumers, AI-native systems remove friction from interactions by allowing subscribers to communicate in natural language rather than navigating disconnected applications and processes. The architecture translates customer intent into system actions, connecting previously isolated systems so requests can be resolved seamlessly. The result is a simpler experience that helps subscribers discover services and outcomes that may not have been visible through traditional workflows.