At Mobile World Congress 2026, the dominant narrative around AI was predictable: more power on-device, larger models, tighter hardware integration. But in a conversation with Rahul Sandil from MediaTek on the sidelines of the event, a different approach emerged. Instead of pushing everything onto the device, the company is thinking more deliberately about where AI should actually run.
That distinction becomes clearer when you move from product pitches to real-world constraints like power, thermals, and cost.
Rethinking Where AI Actually Runs
One of the more telling demos at MediaTek's booth involved a robot that appeared autonomous at first glance. In reality, much of its intelligence was not on the device itself.
"The robot believes it has a brain, but the brain is actually sitting on an edge server," Sandil explained during the interaction.
The setup still included an SoC on the device, but the heavier AI workloads, including large language model processing, were handled externally and accessed over a network connection.
For MediaTek, this is less about showcasing a futuristic demo and more about addressing a practical limitation. Running full-scale AI models locally increases power draw and thermal load, something that becomes harder to manage as models grow larger.
"You don't need to load the entire model on the device," Sandil said, pointing out that doing so would add to battery and energy consumption.
From Device-Centric AI To System-Level Thinking
What MediaTek is outlining here is a shift from device-centric AI to a more distributed model. The device continues to handle latency-sensitive tasks, but larger computations are pushed to the edge, where they can be updated and scaled more easily.
"There is still compute on the device," Sandil noted, "but the larger model can sit on the server and be updated there."
That changes how improvement cycles work. Instead of relying entirely on new hardware for better AI performance, companies can upgrade models on the backend, allowing devices to tap into newer capabilities without a full hardware refresh.
The Network Becomes Critical
This approach is closely tied to advances in connectivity. MediaTek's demonstrations leaned on the idea that next-generation networks will make this kind of distributed compute viable at scale.
If latency and reliability improve to the point where the network becomes effectively invisible to the user, the need to carry full-scale AI models on-device reduces significantly.
That has implications not just for smartphones, but for a wider set of connected devices where power and cost constraints are even tighter.
A More Grounded AI Narrative
What stands out from this interaction is not a single feature or specification, but a shift in framing. While much of the industry continues to position AI as a race to fit more onto the device, MediaTek appears to be taking a more balanced view.
It is not arguing against on-device AI. It is questioning how much of it actually needs to be there.
At a time when AI conversations are increasingly shaped by scale and capability, that perspective feels less about pushing limits and more about managing them.
