Pallavi Mahajan on the melding of AI and advanced connectivity
Pallavi Mahajan has AI on the brain. No, she’s not trying to train a better chatbot. Nor is she exploring the boundaries of artificial general intelligence. Instead, as Nokia’s Chief Technology and AI Officer, Pallavi is laser-focused on the many, many ways AI and networking intersect.
There’s no question AI will reshape network traffic in profound ways, but Pallavi sees connectivity’s relationship with AI as more than just reaction to changing network workloads. She sees a symbiosis. As the AI supercycle progresses the network will become as integral to AI services as GPUs. Advanced connectivity will unlock never-before-seen capabilities in AI, just as AI will become fundamental to new generations of networking.
Pallavi brings a unique perspective to these frenetic shifts in technology. Starting her career in early internet architectures, her expertise expanded into accelerated, high-powered, cloud and edge computing as well as software-defined networking and autonomous networks. At HPE, Pallavi led the pioneering team that created Frontier, the first exascale supercomputer. Later she spearheaded both Intel’s Data Center and AI business and its Network and Edge group building platforms that span the entire compute spectrum and building innovative business models . In short, Pallavi has always been at the critical convergence of computing and networking.
Pallavi recently sat down with Agile Telco Managing Editor George Malim to share her insights into how connectivity will drive the AI supercycle. Here are excerpts from that interview.
How AI is reshaping network traffic
When we open a chatbot on our computer, it’s easy to think of it as another internet application. But when you press “enter” on your query, what happens on the network is radically different. As Pallavi described it to Agile Telco:
“With a chat window, every prompt can trigger coordination across thousands of GPUs and specialized models. Now, the AI workload behaves very differently from web traffic. The first difference is disaggregation, because modern AI inference pipelines are increasingly fragmented across different compute domains, such as racks, pods and data centers and different stages of inference, whether it's reasoning, retrieving, retrieval or agentic coordination, they all now operate across this distributed infrastructure.”
The bottom line, Pallavi says, is our networks have traditionally been designed for a different kind of traffic than that generated by AI – and AI traffic is only increasing at an exponential rate.
“Traditional internet traffic tolerates unpredictability surprisingly well, but AI workloads do not. Once the network becomes unstable, the entire stack above it has to compensate with larger safety margins, more retries and more context handling.”
Video
Connecting intelligence in the AI era
How AI is making us rethink connectivity
Pallavi views AI less as a technology wave and more as a systems problem. She doesn’t describe AI as another application to be carried over existing networks. Rather it as a structural shift in what networks are being asked to do. As she said in the interview:
“The real breakthrough for the AI supercycle isn’t bigger silicon, it’s the network, and that is what turns many machines into systems. As the unit of compute changes, AI workloads themselves are changing and have become far more distributed. They are no longer monolithic, they are fragmented and distributed by design. We are starting to disaggregate inference across multiple dimensions at once.”
The two major stages of inference, prefill and decode, are being separated. Machine learning problems are being divided among different “experts” residing on GPUs located at far-flung points of the network. In an AI grid, the network is not just carrying workloads, it’s part of the infrastructure itself. Pallavi believes the transformation of compute and connectivity must happen simultaneously: “Distributed inference is now a coordinated system rather than a single process and that coordination is the new complexity.”
AI not only creates new traffic patterns, it also demands new control loops and new requirements for determinism, trust and resilience. It changes how people build and operate networks. Pallavi described how these demands are upending the subordinate role the network has traditionally taken to compute:
“Today workload orchestration and network orchestration are starting to happen, but the old model was that the compute is going to decide and the network is going to follow. That model is breaking and AI is fundamentally changing that whole model because two control planes that have to work together continuously are emerging.
...“The future architecture is not just about faster networking. It is about converged orchestration. In this case, the two brains, the two control planes, are coming together and we're building up autonomous networks because compute orchestration without network awareness is just impractical.”
On creating trust in our networks and Nokia’s role in the AI supercycle
As the AI-era progresses, Pallavi sees trust becoming a table-stakes requirement. AI agents won’t just be working behind the scenes to make people’s lives better. These agents will be managing the increasing complexity of the network and the AI workloads traversing it. Automation and closed-loop optimization will be key to advancing connectivity. But for this shift to happen, we will need iron-clad certainty that our networks will behave the way we want them to. Trust and accountability must be designed into our network architectures, Pallavi said.
“The simple truth is if you can't see what the network is doing, you simply cannot trust it. Autonomous infrastructure has to operate as a glass box, not as a black box and it means three things. First of all, it should be inspectable, which means that operators can understand what the system is doing and why. The second is that it is bounded, which means that the system operates within explicit safety limits. Finally, the third is about reversibility, which means that it must always be possible to roll back safely when needed. Glass box autonomy makes networks trustworthy, explainable and reversible.”
Pallavi’s view of the AI era is clear: the network is evolving into something far bigger than a connectivity layer. It is becoming an execution environment, a coordination substrate and a sensing system for intelligent machines. To be certain, more capacity and higher levels of performance are necessary, but these are only starting points.
AI-native architectures will choreograph the increasingly complex dance between RAN, core, access, transport and data center, while embracing the new traffic paradigms of the AI supercyle. People will need to start thinking at software speed. And security, resilience and trust, must be hardwired into the very fabric of connectivity. The intersection of all these ideas happens to be Nokia’s sweet spot.
“For Nokia, we see this as a major opportunity. We operate across data center fabrics, optical transport, autonomous network operations and AI-RAN so, for us, it's not just an opportunity to build high performance connectivity. It’s about creating the programmable control architecture that allows distributed systems to operate coherently and at scale.”
