Inferring Submerged Edge: Cloud Computing and Edge Computing Dual Drive, Changes in US Stock AI Computing Power Investment Logic

Over the past two years, the main trend in US stock AI computing was almost exclusively focused on training: chip manufacturers competed for flagship training cards, cloud providers poured money into expanding data centers, and power and cooling became new scarce resources. However, entering the fourth quarter of 2026, a noteworthy shift emerged in market discussions—the narrative of computing power is moving from centralized training to distributed inference. The boundaries between cloud computing and edge computing are being redefined, and a new investment logic is taking shape. For investors, this means the coordinate system for stock selection needs to expand from whose computing card is strongest to who is closest to the user. ## Why Inference Demand Rewrites the Geographical Distribution of Computing Power The fundamental differences between training and inference determine their differing requirements for infrastructure. Training seeks peak computing power, naturally suited for concentration in large data centers with cheap electricity, cold climates, and ample land. Inference, however, is a service for end-users, requiring completion within millisecond-level user experience constraints for every invocation, making its deployment logic closer to the demand side. The industry generally believes that three factors are driving the shift of inference computing towards the edge: * **Latency Constraints:** Scenarios such as autonomous driving, industrial robotics, AR/VR, and real-time video generation require response times measured in milliseconds; centralized cloud cannot satisfy this for data traveling hundreds of kilometers away. * **Cost Structure:** Sending all inference traffic back to the central cloud implies a triple burden of bandwidth, electricity, and computing costs. Performing preprocessing and filtering at the edge can significantly reduce the cost per invocation. * **Data Sovereignty and Compliance:** Regulations governing data export are tightening in major global economies, and sensitive data in fields like healthcare, finance, and government increasingly requires on-site processing, making edge nodes a necessity for compliance. ## Hyperscalers and Edge Players: Division of Labor, Not Substitution A common misconception is that the rise of edge computing implies the decline of central cloud. The more accurate picture is a layered architecture of coordination between central hubs and peripheral nodes. ### Hyperscale Cloud Providers: Brains and Model Factories Hyperscale providers such as Microsoft, Google, Amazon, and Oracle still control the central aspects of large model training, model hosting, and complex inference. Although the growth rate of their capital expenditure may slow down, their absolute scale remains enormous, and their spending structure is shifting from simply buying cards to building ecosystems, necessitating self-developed chips, network interconnects, and software stack optimizations. For investors, the value anchor for cloud providers is shifting from revenue from compute leasing to the proportion and quality of AI platform revenue. ### CDN and Edge Networks: Natural Carriers for Edge Computing Power CDN providers such as Cloudflare, Akamai, and Fastly have hundreds to thousands of edge nodes globally. These nodes, originally built for content distribution, are ideal carriers for deploying lightweight inference services. When enterprises want to move AI capabilities closer to the user, the existing network of CDN providers implies extremely low marginal expansion costs, which is their core barrier distinguishing them from traditional data center companies. ### Telecommunication Operators and Edge Data Centers Dense base stations and local aggregation rooms are being transformed into edge computing nodes, and the model of operators and cloud vendors co-building edge clouds is gradually maturing. Meanwhile, data center operators such as Equinix and Digital Realty are extending their product lines from large campuses to smaller edge rooms in urban core areas, and their interconnection business benefits from the traffic structure changes brought by compute layering. In other words, the edge is not the opponent of the cloud, but an extension of the cloud; the true pressure point might be intermediate-layer players who lack scale advantages and a close connection to user scenarios.

Edge AI: The Third Pole Easily Overlooked

Closer to users than network edge are the end devices themselves. AI PCs with NPUs, flagship phones, intelligent vehicles with local large model capabilities, and machine vision equipment for real-time inspection in factories collectively form the base of the compute pyramid. Edge inference consumes no cloud resources, does not rely on network connectivity, and naturally satisfies privacy requirements, becoming a new differentiation selling point for device manufacturers. For the chip industry chain, edge AI means increased demand for client and automotive businesses from designers such as Qualcomm, MediaTek, Intel, AMD, and it also means a further diversification of demand structure in areas like storage and advanced packaging. This capital expenditure is borne by consumer electronics and automotive manufacturers, having lower correlation with traditional cloud capital expenditure, offering diversified value for investment portfolios. Worth noting is that the rise of edge and client-side AI is simultaneously changing the demand curves for network and supporting components: fiber optics for city domains, cooling for edge rooms, miniaturized power supplies, and server forms optimized for inference are all spurring new niche markets.

Investment Themes and Risk Warnings

The trend of convergence between cloud computing and edge computing can be summarized into three investment themes:

• Mid-level: Hyper-scale cloud vendors with self-developed chip and AI platform capabilities, focusing on marginal changes in AI revenue ratio and Return on Invested Capital (ROIC).
• Edge Layer: CDN vendors, edge data center operators, and telecommunications infrastructure companies, focusing on the pace of edge AI service orders and revenue per node.
• Client-side Layer: End-device NPUs, storage, and advanced packaging suppliers, focusing on the upward slope of AI PC and intelligent vehicle penetration rates.

Risks cannot be ignored: First, edge compute demand is still in the early stages; if killer applications do not appear soon, the valuation premium of related companies may be reversed. Second, inference technology is rapidly iterating; model miniaturization and distillation techniques may cause some inference workloads to concentrate back in central clouds, suppressing the edge narrative. Third, fluctuations in macroeconomic interest rates and the overall valuation of the technology sector will still lead to phased pullbacks in high-multiple growth stocks. Overall, the story of AI compute is shifting from building more to placing more cleverly. Cloud provides scale, edge computing provides latency and compliance, and client-side provides privacy and personalization, jointly forming a complete inference supply system. For investors focusing on US stocks in the AI compute theme, rather than betting on extreme elasticity in a single arena, it is better to follow the layered map of mid-level, edge, and client-side, seeking companies with genuine revenue realization capabilities in each segment; this might be the more stable way to find gold in the next phase.