Nokia sees gen AI enabling conversational networks in three to five years
Nokia recently laid out its vision for a future network architecture that’s all about creating a network-cloud continuum that spans devices, edge and regional infrastructure, and the network core—and with both more classical AI and geg AI ultimately present throughout the platform. Nokia’s CTO for Cloud and Network Services Jitin Bhandari told attendees to the Telco AI Forum 2.0 (available on demand here) that the combo of cloud and AI is necessary for a new age of consumer, enterprise and industrial services. But, “The acceptance and the acceleration of cloud and AI is going to be a challenge. We have to understand that.”
Bhandari explained that Nokia estimates that CSPs “have a unique opportunity to tap into an about additional $1.1 trillion market…with new opportunities in the spaces of many, many verticals.” But that requires networks capable of serving those verticals with agility and scale. “It’s easier said than done. There are many building blocks underneath that lead up to these foundational pieces.”
He said that CSPs have a “very unique genome” suited to building and operating large scale networks. This core competency allows them to “pivot off into this cloud continuum and build up cloud practices, build up security practices, build up AI practices…if the networks are foundationally built right…Needless to say, automation, AI, cloud, security, these are technology paradigms that are front and center to this architecture.”
CSPs are no stranger to AI. In fact, Bhandari said, Nokia has delivered 80-plus use cases with more than 300 operators that employ ML-based techniques. The next step, he said, is conversational AI and agentic AI wherein humans use natural language to communicate with the network, and autonomous systems are allowed to take particular actions without manual intervention. Bhandari looked at the blending of structured and unstructured data into multi-modal gen AI models which “makes the AI offering much more powerful. Networks are going to be conversational in three to five years. That’s what we strongly believe.” So how do CSPs get from the present to that future in a half-decade?
Bhandari suggested three areas of focus:
- Use cases—“It’s all about understanding the persona of the human that’s associated with the cycle of the network…Think about the operations and think about the business outcome, and define the use case crystal clear.”
- Data—data constructs, governance, observability, privacy and sources. “When you play with unstructured data, this is one technology where you can do a [proof of concept] or a trial in two weeks but when you want to build a system for scale, it takes quite a bit.”
- Model selection—there are a variety of closed source and open source LLMs, and other model modalities, available to use or buy and then augment. And larger models aren’t necessarily better; smaller models infused with domain-specific data can often deliver faster time-to-value. “Start thinking…how do you build up model farms, model gardens?”
More on this idea of building out of model farm or garden: Bhandari emphasized the wide range of AI techniques—knowledge graphs, low-rank adaptation, model tuning, retrieval augmented generation, vectorization of unstructured data, etc…—and the fundamental data science skillsets that are needed to do those things. “It’s less about the foundation model, more about the operations techniques around it,” he said.
