Qualcomm believes the 6G era will be defined not by speed, but by the deep integration of artificial intelligence into network architecture, according to a top executive. Durga Malladi, Qualcomm's Executive Vice President and General Manager of Edge Solutions and Data Center Business, shared his vision during a media interview at the company's global 6G analyst and media day in San Diego. He outlined how 6G terminals may evolve beyond smartphones into a new device category, designed from the ground up for AI workloads.
Malladi stated that the true watershed for 6G is not network speed but the fact that AI is being written into the foundational logic of the network for the first time. He described 6G as an "AI-native technology," meaning networks are designed from the outset to incorporate high-performance computing processors and leverage AI for autonomous operation. He noted that many operators are already investing in this area, beginning the transition of their existing networks toward an AI-native architecture.
When asked about the most fundamental difference between 6G and 5G, Malladi explained that the distinction lies not in bandwidth improvements, but in the ubiquitous native integration of AI. He acknowledged that physical layer technical standards are still being developed, with definitive answers expected once the 6G standard is finalized in 2028. He added that an AI-native network is built on high-performance computing units, including CPUs and AI accelerators, integrating numerous air-cooled servers and accelerator cards throughout the mobile network, from radio units to the core network. The key characteristics include AI inference capabilities alongside the ability to run wireless protocols on standard processors.
Regarding terminal form factors, Malladi believes the 6G era will give rise to a new class of devices he calls "intelligent agent AI devices." This trend is already emerging in China, where he specifically cited AI-powered phones as an example. In his view, these new AI devices are not limited to smartphones, encompassing sensors, wearables, and other connected devices, some of which may not even have displays. He mentioned that major manufacturers are also developing agent-first terminals, suggesting this differentiation will drive the entire industry forward.
Malladi observed that 6G will continue the trajectory of 5G in terms of connectivity performance, such as faster deployment, but speed is not the critical factor. The more important difference lies in the widespread application of AI in both networks and devices. He also noted that cloud AI companies are increasing their hardware investments, potentially positioning themselves as a new generation of OEMs. Regarding how AI and 6G will deeply integrate, he admitted the implementation path is still being explored.
Qualcomm is currently working closely with operators in the United States, China, Japan, Korea, and Europe. In China, the company is participating in pre-commercial technology trials, and Qualcomm is planning a 6G network demonstration for the 2028 Los Angeles Olympics. Malladi believes 6G will drive a substantial shift in operator business models. He explained that if operators integrate AI capabilities and substantial processing power into their 6G networks, they will likely want to offer new service models, such as computing services and Token services, to recover their investments and generate revenue. He again referenced China, citing a major Chinese telecom operator as an example of this evolving model, as it has already added computing power as a new service offering alongside traditional data services.
Regarding Qualcomm's approach to the competitive data center market, Malladi stated the company's primary entry point is through AI accelerators and NPU solutions specifically optimized for generative AI applications. The company also leverages high-bandwidth computing and custom silicon services to meet data center demands. He mentioned that data centers use large-scale, liquid-cooled and air-cooled servers, which differ from smartphone architectures. For example, Qualcomm's AI250 solution currently offers 160 kilowatts of power, with expectations to reach hundreds of kilowatts in the future.
In custom silicon, Qualcomm provides its own IP while also integrating customer-provided IP to meet specific needs. Malladi explained that these product lines reuse underlying technology through a scalable architecture while being customized for different customer requirements. On the topic of high-bandwidth memory, he explained why Qualcomm chose high-bandwidth computing (HBC) over the mainstream HBM approach. HBC adopts a near-memory computing design philosophy, placing compute logic as close to memory as possible, which reduces power consumption and improves bandwidth utilization compared to HBM. He clarified that HBC does not alter the functionality or pricing mechanisms of memory itself, nor does it fundamentally change the role of the memory industry. Qualcomm is working closely with memory manufacturers to optimize the distance between computation and memory.
On the question of distributing inference workloads between the edge and the cloud, Malladi said the decision rests not with humans but with AI agents themselves. He explained that agents decide at each moment whether a task should run on the terminal, at the edge, or in the cloud. Users only perceive the final result and need not be concerned with where it occurs. He added that most inference currently happens in the cloud, but as technology evolves, more tasks will shift to the edge and base stations, with the exact proportion varying by scenario for classical AI models versus generative models. He emphasized that Qualcomm's role is to introduce and demonstrate the value of this concept to operators, not to force acceptance. Some operators will adopt it quickly, while others may maintain the status quo, with final decisions based on their own business logic.
Malladi also addressed the "continuum of computing" concept that Qualcomm frequently highlights. He described it as the idea that computing is no longer confined to a single location but extends seamlessly from terminals to data centers. Sensing capabilities were also discussed, with Malladi noting that sensing technology enables the network to comprehensively perceive internal information and the environment, detecting and analyzing pedestrians, vehicles, traffic lights, drones, and other elements in real time. This facilitates the construction of autonomous networks that can self-adjust parameters, similar to ADAS systems in automobiles.
HBC represents Qualcomm's direct challenge to the existing HBM architecture dominant in AI data centers. Malladi explained that HBM places memory and computation on separate sides with a high-speed interface for data transfer, while HBC uses near-memory computing, placing some compute logic directly beneath the memory. This proximity yields two benefits: significantly improved effective bandwidth and reduced power consumption. He noted that HBM inevitably leads to higher power consumption, which is the core problem HBC aims to solve. As the third of four data center product lines, HBC primarily targets the decode stage, which is the most demanding bottleneck in large model inference. (Qualcomm's four data center product lines include CPU, AI accelerators, HBC, and custom silicon/ASIC services.)
Qualcomm is currently the only HBC supplier, making its key task the persuasion of hyperscale cloud providers to accept this approach. Malladi said these cloud giants are naturally attracted to the technology and have proactively approached Qualcomm. Meanwhile, the company is in discussions with all memory manufacturers, since DRAM stacking ultimately relies on them, and these communications are going smoothly. He stressed that HBC is not a bundled product; if a hyperscaler already has its own CPU and accelerator cards, it can purchase HBC independently and integrate its own compute logic as a standalone component. As for whether HBC will change the fundamental structure of the memory industry, Malladi gave a direct answer: no. HBC simply brings computation and memory closer together without altering memory architecture itself, so Qualcomm and memory makers are partners, not replacements. Regarding physical layer innovation, he stated that Qualcomm has invested heavily in probabilistic shaping and geometric shaping to improve spectral efficiency, while also increasing investment in massive MIMO radio frequency and network-side AI technologies.