Over the past decade, edge computing has been a recurring buzzword at various conferences.

In the early days, the industry focused on defining 'what is edge computing' and 'why it is needed,' leading to debates over terms like far edge, near edge, and network edge. Later, edge computing became intertwined with 5G, industrial internet, and connected vehicles, with core keywords shifting to 'real-time performance' and 'local decision-making.' Operators and equipment vendors began promoting MEC (Multi-Access Edge Computing) architectures. As concepts clarified, AI inference started migrating to the edge, enabling cameras, robots, and industrial devices to analyze data locally and respond in real time rather than merely collecting it. Now, with the rise of generative AI and agents, industry discussions have shifted from 'computing power sink (decentralization of computing power)' to 'distributed intelligent collaboration.' Many edge computing visions that once remained theoretical are now accelerating into real-world industrial scenarios.

Recently, at an overseas symposium, experts from JLL, Intel, Ericsson, Qualcomm, and American Tower Corporation discussed the latest trends in edge computing. Participants spanned multiple sectors of the industrial chain, including real estate, chips, communication equipment, and tower infrastructure. Many of their insights were thought-provoking, and this article aims to organize and share them.

The 'Belated' Killer Application: AI Inference

Rather than attempting to establish a single standard definition for 'edge,' participants increasingly embraced a 'continuum' consensus—a flexible, programmable execution environment spanning central clouds, regional edges, local edges, and even enterprise edges. In this system, workloads are dynamically deployed to different locations based on latency, privacy, security, and cost requirements.

One of the most notable viewpoint (viewpoints) from the discussion was that AI inference is emerging as the true 'killer application' for edge computing.

Jim Poole of American Tower Corporation succinctly summarized the past decade of MEC development with a vivid metaphor: 'MEC was like searching for a hammer while holding a nail.' Multi-Access Edge Computing (MEC) is a network architecture that provides cloud computing capabilities and IT service environments at the network edge, aiming to reduce latency, ensure efficient network operations and service delivery, and improve customer experience.

From the definition of MEC, it is clear that the industry long anticipated the need for a distributed computing layer, prompting operators and infrastructure vendors to deploy numerous edge nodes in advance. However, the problem lay in the absence of business scenarios that truly required these nodes. In other words, infrastructure outpaced demand. This observation also explains why edge computing remained more hype than reality for years. While industrial internet, VR/AR, and connected vehicles were all considered key directions for edge computing, none generated sustained, large-scale computing demand.

Today, the difference lies in a fundamental shift on the demand side. With the proliferation of generative AI, regardless of the answers users seek, they must provide requests and data sources through uplink connections. Raw data such as high-definition images, audio, and video streams generated locally must now be uploaded to the cloud in real time for processing, leading to a dramatic increase in uplink data volumes over the past year. Dr. Koymen of Qualcomm noted that user behavior is shifting from downlink-centric video consumption to uplink-centric AI-generated traffic, with agentic data expected to surpass human-generated data in the coming years. Joe Constantine of Ericsson reinforced this point by citing data from the Ericsson Mobility Report: global data traffic is projected to triple by 2029, while uplink traffic will grow tenfold by 2035.

This new 'upload-infer-respond' model imposes unprecedented demands on network latency and bandwidth—precisely where edge computing excels. Sean Farney of JLL declared, 'Edge AI inference is making infrastructure sexy again.' After two decades of pursuit, the industry has finally arrived at its true killer application: AI inference. It possesses two critical attributes: sufficiently high computational density and extreme sensitivity to latency. Together, these factors 'force' computing power to decentralize from centralized cloud data centers.

AI Is Rewriting the Data Center Paradigm

What, then, will a true AI-era edge node look like? Poole shared striking data: over the past 25 years, approximately 95% of global data centers were designed around a power density of 5–10kW per rack. Today, next-generation AI systems demand 150–200kW per rack, with Google even showcasing a 1MW-per-rack configuration.

This is no longer a problem solvable through minor airflow optimizations. Two direct consequences have emerged. First, 'enterprise self-built data centers' are rapidly becoming infeasible. Poole noted that for two decades, the data center industry's primary competitor was enterprises' in-house server rooms. However, with the soaring complexity and power density of AI infrastructure, self-built models are increasingly untenable. 'You can no longer solve the problem by simply building your own data center,' he said.

Second, liquid cooling is transitioning from an 'advanced solution' to an industry standard, while on-site power generation is becoming a regulatory requirement in some U.S. states.

Simultaneously, the geographical distribution of computing infrastructure is undergoing profound changes. While most North American computing resources remain concentrated in roughly 15 core metropolitan areas today, Poole predicted that the industry will rapidly expand to 30–50 secondary and tertiary markets within a timeframe far shorter than the past 25 years.

The truly unresolved question is whether future edge infrastructure will trend toward 'centralization' or 'decentralization': Will the future hold 300 national 10MW-class data facilities, or 2,000 60kW edge cabinets deployed beside every communication tower? Tower companies clearly favor the latter.

As a representative of chip vendors, Qualcomm offered an additional perspective: not all AI inference tasks require GPUs. Koymen argued that while GPUs excel at model training, their cost and power consumption are prohibitive for inference. Instead, NPUs optimized for inference and deployed on end devices and far-edge locations are better suited for lightweight inference tasks across the edge continuum.

In a sense, AI is restoring data centers to their 'heavy industry' roots. This also means that competition in edge computing is shifting from software capabilities to energy, real estate, and infrastructure capabilities—namely, who can secure land, power, cooling, and deployment resources faster. This explains the presence of real estate and tower infrastructure firms in the discussion.

2028–2029: What Lies Ahead?

Facing impending transformations, experts offered concrete predictions for the 2028–2029 industry landscape.

Qualcomm's Koymen tied his forecast to the company's 6G roadmap: pre-commercial devices will emerge in 2028, with global commercialization synchronized with operators by 2029. By then, edge infrastructure will support AI Recall, 'what-you-see-is-what-you-get' AR glasses, and distributed robotic computing.

Ericsson's Constantine provided three more quantifiable judgments: first, by 2029, 75% of global data traffic will run on 5G networks; second, the industry will stop debating 'what the edge is' and instead argue over service-level agreements (SLAs) and TM Forum Level 4/5 automation—a sign of maturity; third, sustainability pressures will become the primary design constraint for data centers.

Intel's Agarwal focused on industrial implementation, predicting that by 2028, industry conferences will feature retailers, mining companies, and port operators discussing actual ROI from deployments rather than vendors discussing architectures. He warned against repeating the mistakes of private wireless networks, which were built but failed to produce widespread success stories.

Farney offered the most far-reaching prediction: humanoid robots will begin appearing in data center operations to address severe labor shortages. This concept of 'physical AI' is becoming reality—NVIDIA and T-Mobile have announced a partnership to deploy AI-RAN infrastructure at the 5G network edge, enabling AI agents to perceive and respond in real time at urban intersections and industrial facilities while drastically reducing terminal device requirements through edge computing power.

Epilogue

However, beyond the technology itself, the discussion ultimately highlighted several more pragmatic issues: power, talent, and data.

First is power: AI inference is causing computing demand to explode, but global grid construction lags far behind. Poole bluntly stated that the U.S. grid was not designed for such localized, high-density loads. Second is talent: JLL alone currently has over a thousand unfilled data center-related positions. Finally, there is data: Constantine argued that future winners may not be companies with the best models but those with the highest-quality data systems. As model capabilities converge, data quality, structure, and governance will likely become the true moats in AI competition.

Taken together, a clear picture emerges: edge computing has moved beyond theoretical 'why' discussions into the 'how-to-build-it' engineering phase.

References: Why the edge finally has its killer use case – RCRWireless; Times Have Changed: 'Large Uplink' Becomes the Focus of Communication Network Upgrades – moomoo; What Is Multi-Access Edge Computing (MEC)? – Redhat