At a 6G media day in San Diego, Qualcomm executive vice president and general manager of edge solutions and data center business Durga Malladi set the tone for the next generation of mobile networks: the dividing line isn't speed. His exact words were that "the differentiator for 6G isn't higher bandwidth and data rates — it's the native, everywhere integration of AI." Per Qualcomm's own timeline, the 6G physical-layer technical standard is still being drafted, with final ratification expected in 2028.
A device category that doesn't have a name yet
The concept Malladi put forward is "agentic AI devices." His definition is far broader than smartphones: it covers sensors, wearables and a whole range of connected endpoints, some of which may not even have a display. The example he cited comes from the Chinese market — the Doubao AI phone.
Pull out the phrase "no display" and what Qualcomm is describing is a class of devices built around voice and sensing as the primary interface, with computation split between the device and the network rather than sitting entirely on the device. This kind of product has come in several waves over the past decade, from smart speakers to various AI wearables, and none of them have had a high survival rate. What's different this time, in Qualcomm's telling, is the premise: the network itself will now shoulder part of the inference load, so the device no longer has to carry it alone.
Rewriting the carrier revenue line
The second part of his argument has a more direct commercial meaning. Malladi believes carriers' business model will shift from selling data to "compute as a service, tokens as a service." The example he gave was China Telecom, which he said has already added compute services on top of its traditional data services.
The weight of that statement is in the billing unit. Selling bandwidth is priced by the gigabyte, and per-unit prices have been sliding for years, which is why carrier ARPU has been under pressure for a decade. Pricing by tokens instead measures the number of inference calls, which tracks directly with how intensively a user relies on AI — a completely different ceiling. The difficulty sits in the same place: carriers don't own models of their own, so to get into the token business they first need inference clusters, and their competitors there are the cloud providers.
Qualcomm's own bet
Behind this narrative is Qualcomm's data center business. The company now has four product lines in place — CPUs, AI accelerators, HBC (high-bandwidth compute), and custom silicon/ASICs — with a target of more than $15 billion in data center revenue by fiscal year 2029. Its AI250 solution runs at 160 kilowatts per rack, and Qualcomm expects that to climb into the hundreds of kilowatts in the future.
On the technology side, Qualcomm didn't follow the industry toward HBM; it's using its own HBC instead, arguing that "HBC's near-memory-compute design cuts power consumption and improves bandwidth utilization more effectively than HBM." By its own account, cloud providers and memory makers have already signaled support. Modular, the AI infrastructure company Qualcomm acquired earlier, fills in the software-stack side of that pitch.
A timing gap worth noting
There's one thing worth flagging. The 6G standard is due to be finalized in 2028, and going by the pace of previous generations, large-scale commercial deployment wouldn't arrive until around 2030. "Tokens as a service," meanwhile, is something China Telecom is already doing now — it has no dependency on 6G at all.
Tying the two together works in the chip maker's favor: it makes that $15 billion fiscal-2029 target look like it sits on a path leading toward a standard. For carriers, though, whether they can actually collect token revenue depends on whether they can build sellable inference capacity within the next four or five years — and that has little to do with the base-station upgrade cycle.
Sources: Qualcomm 6G Media Day interviews, Sina Technology, CocoLoop; verified against Malladi's title, the 6G standard timeline, the data center revenue target, and the AI250 power figures.