Why has edge computing finally found its killer application?

Over the past decade, edge computing has been a frequently discussed hot topic at various conferences.

In the early days, the industry's focus was on "what is edge computing" and "why is it needed." The industry was caught in debates over definitions such as far edge, near edge, and network edge. Later, edge computing began to be associated with 5G, the industrial Internet, and the Internet of Vehicles. The core keywords became "real - time" and "local decision - making." Operators and equipment manufacturers started to promote the MEC (Multi - Access Edge Computing) architecture. Subsequently, as the concept became clearer, AI inference began to move to the edge side. Cameras, robots, and industrial equipment not only collect data but can also perform local analysis and real - time response. Nowadays, with the rise of generative AI and Agents, the industry's discussion focus has shifted from "computing power sinking" to "distributed intelligent collaboration" - many edge computing concepts that used to stay at the theoretical level are being rapidly implemented in real industrial scenarios.

Recently, at a special symposium held overseas, multiple experts from JLL, Intel, Ericsson, Qualcomm, and American Tower Corporation discussed the latest trends in edge computing. The participants covered multiple links in the industrial chain, including real estate, chips, communication equipment, and tower infrastructure. Many viewpoints were quite inspiring, and this article sorts them out and shares them.

The "Late - Arriving" Killer Application: AI Inference

The participants did not attempt to give a single standard definition of "edge" but gradually reached a consensus on a "continuum," that is, a flexible and programmable execution environment spanning the central cloud, regional edge, local edge, and even enterprise edge. In this system, workloads are dynamically deployed to different locations according to requirements for latency, privacy, security, and cost.

One of the most notable viewpoints in this discussion is: AI Inference is becoming the real "killer application" of edge computing.

Jim Poole from American Tower Corporation summarized the development of MEC over the past decade with a rather vivid statement: "MEC is like looking for a hammer while holding a nail." The so - called Multi - Access Edge Computing (MEC) is a network architecture that provides cloud computing functions and IT service environments at the network edge. Its goal is to reduce latency, ensure efficient network operation and service delivery, and improve the customer experience.

From the definition of MEC, the industry actually predicted early on that a distributed computing layer would be needed in the future. Therefore, operators and infrastructure manufacturers pre - arranged a large number of edge nodes. However, the problem was that there were no real business scenarios that required these nodes at that time. In other words, the infrastructure was ahead of the demand. This judgment also explains why edge computing has been "much ado about nothing" for the past few years. Whether it is the industrial Internet, VR/AR, or the Internet of Vehicles, although they are all regarded as important directions for edge computing, these scenarios have not formed a large - scale and sustainable computing power demand.

The difference now lies in a fundamental change on the demand side. With the popularization of generative AI, no matter what answers users need, they need to upload requests and data sources. The original data such as high - definition images, audio, and video streams generated locally must be uploaded to the cloud in real - time for processing, resulting in a significant increase in upstream data volume in the past year. Dr. Koymen from Qualcomm said that user behavior is shifting from video consumption mainly on the downstream link to AI - generated traffic centered on the upstream link. Agentic data will exceed human - generated data in the next few years. Joe Constantine from Ericsson cited data from the "Ericsson Mobility Market Report" to further confirm this: Global data traffic will triple by 2029, and by 2035, upstream link traffic will increase tenfold.

This new model characterized by "upload - inference - response" places unprecedented demands on network latency and bandwidth, which is exactly where edge computing comes in. Sean Farney from JLL made an assertion: "Edge AI inference is making the infrastructure field sexy again." After chasing this goal for 20 years, the industry has finally welcomed the real killer application - AI inference. It has two key characteristics: high enough computing density and high sensitivity to latency. These two factors together "force" the computing power to spread from the centralized cloud data center.

AI is Forcing the Data Center System to be "Rewritten"

So, what exactly will a real edge node for the AI era look like? Poole presented a set of extremely impactful data: In the past 25 years, about 95% of global data centers were basically designed with a power density of 5 - 10kW per cabinet. Now, the power density of the new - generation AI systems has reached 150 to 200kW per cabinet, and Google has even demonstrated a configuration of 1MW per cabinet.

This is not a problem that can be solved by simply optimizing the air duct design. This directly brings two changes - First, "enterprise - built data centers" are quickly losing their feasibility. Poole said that in the past two decades, the biggest competitor in the data center industry has actually been the in - house computer rooms built by enterprises. However, with the rapid increase in the complexity and power density of AI infrastructure, the self - built model is becoming increasingly difficult to establish. "You can no longer solve the problem by building your own data center as you did in the past."

Second, liquid cooling is changing from an "advanced solution" to an industry standard, and on - site power generation is even gradually evolving from an "optional item" to a regulatory requirement in some states in the United States.

At the same time, the geographical distribution of computing power infrastructure is also undergoing great changes. Currently, most computing power resources in North America are still concentrated in about 15 core metropolitan areas. However, Poole predicts that in a much shorter development cycle than the past 25 years, the industry will quickly expand to 30 to 50 second - and third - tier markets.

The real unresolved question is whether the future edge infrastructure will move towards "centralization" or "discretization": Will there be 300 data facilities at the 10MW level distributed across the country, or will 2000 60kW edge cabinets be deployed beside each communication tower? Tower companies obviously prefer the latter.

As a representative of chip manufacturers, Qualcomm's view adds another perspective: Not all AI inference tasks must rely on GPUs. Koymen believes that GPUs are very suitable for model training, but in inference scenarios, their cost and power consumption are too high. In contrast, NPUs deployed on terminal devices and the far - edge side, which are specifically optimized for inference, are more suitable for undertaking lightweight inference tasks in the edge continuum.

In a sense, AI is bringing the data center back to its "heavy industry" nature. This also means that the competition in edge computing is gradually shifting from software capabilities to energy, real estate, and infrastructure capabilities, that is, who can obtain land, power, heat dissipation, and deployment resources faster - this is why real estate and tower infrastructure companies are involved in this discussion.

2028 - 2029: Where Will the Industry Go?

Facing the upcoming changes, experts also gave specific predictions for the industry landscape in 2028 - 2029.

Koymen from Qualcomm linked his prediction to the 6G roadmap: Pre - commercial devices will appear in 2028, and commercialization will be achieved in synchronization with global operators in 2029. By then, edge infrastructure will support application scenarios such as AI Recall, "what you see is what you get" AR glasses, and distributed computing for robots.

Constantine from Ericsson gave three more quantitative judgments: First, by 2029, 75% of global data traffic will run on 5G networks; second, the industry will no longer debate "what is edge" but will shift to debating service - level agreements (SLAs) and TM Forum Level 4/5 automation - this is a sign of the industry's transition from a wild - west stage to a mature stage; third, the pressure on the sustainable development of data centers will become the primary design constraint.

Agarwal from Intel's prediction is more focused on industrial implementation. He believes that by 2028, similar industry seminars will look completely different. Instead of equipment manufacturers discussing architectures on the stage, retailers, mining companies, and port operators will share their actual return on investment from deployment. He warned the industry to avoid repeating the mistakes of private wireless networks: The networks were built, but successful cases never really emerged.

Farney's prediction is the most long - term: Humanoid robots will start to appear in data center operations to help fill the huge labor gap. The concept of "physical AI" is becoming a reality - NVIDIA and T - Mobile have announced a cooperation to deploy AI - RAN infrastructure at the edge of the 5G network, enabling AI agents to sense and respond in real - time at urban intersections and industrial facilities, while significantly reducing the requirements for terminal devices with the help of edge computing power.

Conclusion

However, compared with the technology itself, this discussion finally pointed out several more practical problems: power, talent, and data.

First is power. AI inference is causing a sharp explosion in computing power demand, but the global power grid construction speed is far behind. Poole even said bluntly that the US power grid is not designed for such local high - density loads. Second is talent. Just JLL alone currently has thousands of data - center - related job vacancies. Finally, it is data. Constantine put forward a notable judgment - the companies that will truly succeed in the future may not be those with the best models but those with the highest - quality data systems. Because as the model capabilities gradually converge, data quality, data structure, and data governance capabilities are likely to be the real moats in future AI competition.

Overall, a clear picture emerges: Edge computing has passed the theoretical demonstration stage of "why it is needed" and entered the engineering implementation stage of "how to build it well."

Reference materials: Why the edge finally has its killer use case — RCRWireless The times have changed, and "large upstream" has become the focus of communication network upgrades — moomoo What is Multi - Access Edge Computing (MEC)? — Redhat

This article is from the WeChat official account "Internet of Things Think Tank" (ID: iot101), author: Sophia. It is published by 36Kr with authorization.